# Graph RAG アプリケーションの構築 ## 概要 TiDB ベクトル検索を使用すると、データの文字値ではなく、データの意味を検索できます。また、生成 AI を搭載したアプリケーションを構築したり、テキスト、画像、ビデオ、オーディオ、またはあらゆる種類のデータに対してセマンティック検索や類似性検索を実装したりすることもできます。ベクトル検索では、埋め込み空間の距離を使用して類似性を表します。ベクトル埋め込みを TiDB に保存した後、関連データを検索するには、ベクトル埋め込みの最近傍を検索するだけです。ベクトル検索機能を備えたリレーショナル データベースである TiDB では、データとそのベクトル埋め込みを 1 つのデータベースにまとめて保存することもできます。同じテーブルに保存することも、異なるテーブルに分けてナレッジ グラフを作成し、後でベクトル検索クエリと JOIN を使用して取得することもできます。 ![52%](https://lab-static.pingcap.com/quick-demo/TiDBAsVectorStoreForRAG_aa_001_20240610.png) ### 所要時間 + この演習は約 90 分程度かかります。 ### タスク + TiDB Cloud Serverless クラスターに接続し、MySQL クライアントからベクトル データを操作します。 + TiDB Cloud Serverless でベクトル関数を使用します。 + 高速な近似最近傍検索のために、HNSW (Hierarchical Navigable Small World) インデックスを作成して使用します。 + TiDB をベクトル ストアとして活用することで、AI アプリケーションの作成が強化され、LLM のナレッジ ベースの境界を超えた問い合わせに対応できるようになります。これらのアプリケーションは、検索拡張生成 (RAG) アプリケーションとも呼ばれます。 + ナレッジグラフアプローチを使用して、RAG アプリケーションを強化します。 ### 前提条件 + このハンズオン演習では、独自の OpenAI API キーを持参する必要があります。 > [!TIP] > 演習環境を起動するには、**開始**をクリックしてください。 > 環境の準備が整ったら、下部のツールバーで **次のモジュール** を選択して **モジュール 1** に進みます。 > そして、ステップバイステップの演習ガイドに従って、ハンズオン環境に接続し、タスクを開始してください。 ## モジュール 1: TiDB ベクトル データ型とベクトル関数の概要 ![48%](https://lab-static.pingcap.com/quick-demo/TiDBVectors_aa_001_20240610.png) ### 手順 1. Linux マシンにログインする + ここをクリック - [ターミナル 1](ssh://monitor1-ssh#1) で、ハンズオン演習用に割り当てられた仮想マシンに接続された新しいターミナルを開きます。 + ログイン後に表示されるバナー： ```output Last login: Mon Jun 3 01:38:55 2024 from ... [rocky@ip-10-90-1-234 ~]$ ``` ### 手順 2. TiDB Cloud Serverless データベースに接続する + 以下の操作はリモート Linux 上で行います。以下のコマンドを使用して TiDB Cloud Serverless に接続します。(MySQL クライアントを終了する場合は `EXIT` を使用します。) ```command mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt ``` ```output mysql: [Warning] Using a password on the command line interface can be insecure. Welcome to the MySQL monitor. Commands end with ; or \g. Your MySQL connection id is 2168969 Server version: 5.7.28-TiDB-Serverless TiDB Server (Apache License 2.0) Community Edition, MySQL 5.7 compatible Copyright (c) 2000, 2024, Oracle and/or its affiliates. Oracle is a registered trademark of Oracle Corporation and/or its affiliates. Other names may be trademarks of their respective owners. Type 'help;' or '\h' for help. Type '\c' to clear the current input statement. mysql> ``` ### 手順 3. ベクトルを作成してクエリを実行する + ベクトルを挿入し、ベクトルをクエリします。 + 新しいスキーマを作成し、新しいスキーマに切り替えます。 ```sql DROP DATABASE IF EXISTS vs_example; CREATE DATABASE vs_example; USE vs_example; ``` ```output mysql> DROP DATABASE IF EXISTS vs_example; Query OK, 0 rows affected (0.00 sec) mysql> CREATE DATABASE vs_example; Query OK, 0 rows affected (0.12 sec) mysql> USE vs_example; Database changed ``` + `VECTOR(3)` を使用して 3 次元ベクトル列を持つテーブルを作成し、そこに 3 行を挿入します。 ```sql CREATE TABLE vector_table ( id BIGINT AUTO_RANDOM PRIMARY KEY, doc TEXT, embedding VECTOR(3)); ``` ```output Query OK, 0 rows affected, 1 warning (0.13 sec) ``` ```sql INSERT INTO vector_table (doc, embedding) VALUES ('apple', '[1,1,1]'), ('banana', '[1,1,2]'), ('cat', '[2, 2, 3]'); ``` ```output Query OK, 3 rows affected (0.08 sec) Records: 3 Duplicates: 0 Warnings: 0 ``` ```sql SELECT doc, embedding FROM vector_table ORDER BY embedding; ``` ```output +--------+-----------+ | doc | embedding | +--------+-----------+ | apple | [1,1,1] | | banana | [1,1,2] | | cat | [2,2,3] | +--------+-----------+ 3 rows in set (0.00 sec) ``` ### 手順 4. ベクトルで関数を使用する + ベクトル場と特定のベクトル `[3,2,1]` 間のベクトル余弦距離を取得します。 ```sql SELECT doc, embedding, VEC_COSINE_DISTANCE(embedding, '[3,2,1]') AS distance_to_321 FROM vector_table ORDER BY distance_to_321; ``` ```output +--------+-----------+---------------------+ | doc | embedding | distance_to_321 | +--------+-----------+---------------------+ | apple | [1,1,1] | 0.07417990022744858 | | cat | [2,2,3] | 0.15733515938218623 | | banana | [1,1,2] | 0.23623738417402662 | +--------+-----------+---------------------+ 3 rows in set (0.01 sec) ``` + ベクトル データ型でサポートされているベクトル関数を使用します。 + ベクトル列のコサイン距離、L1距離（マンハッタン距離）、L2距離（2乗ユークリッド距離）、および負の内積を `[3,2,1]` に対してテストします。 ```sql SELECT embedding, VEC_COSINE_DISTANCE(embedding, '[3,2,1]') AS cos_to_321, VEC_L1_DISTANCE(embedding, '[3,2,1]') AS l1_to_321, VEC_L2_DISTANCE(embedding, '[3,2,1]') AS l2_to_321, VEC_NEGATIVE_INNER_PRODUCT(embedding, '[3,2,1]') AS nip_to_321 FROM vector_table; ``` ```output +-----------+---------------------+-----------+-------------------+------------+ | embedding | cos_to_321 | l1_to_321 | l2_to_321 | nip_to_321 | +-----------+---------------------+-----------+-------------------+------------+ | [1,1,1] | 0.07417990022744858 | 3 | 2.23606797749979 | -6 | | [1,1,2] | 0.23623738417402662 | 4 | 2.449489742783178 | -7 | | [2,2,3] | 0.15733515938218623 | 3 | 2.23606797749979 | -13 | +-----------+---------------------+-----------+-------------------+------------+ 3 rows in set (0.01 sec) ``` + 次元数、ベクトルとテキスト間の変換、ユークリッドノルムなどのその他のサポートされているベクトル関数。 ```sql SELECT embedding, VEC_DIMS(embedding) AS "#_of_dim", VEC_AS_TEXT(embedding) AS to_text, VEC_FROM_TEXT(VEC_AS_TEXT(embedding)) AS embedding_clone, VEC_L2_NORM(embedding) AS l2_norm FROM vector_table ORDER BY embedding; ``` ```output +-----------+----------+---------+-----------------+--------------------+ | embedding | #_of_dim | to_text | embedding_clone | l2_norm | +-----------+----------+---------+-----------------+--------------------+ | [1,1,1] | 3 | [1,1,1] | [1,1,1] | 1.7320508075688772 | | [1,1,2] | 3 | [1,1,2] | [1,1,2] | 2.449489742783178 | | [2,2,3] | 3 | [2,2,3] | [2,2,3] | 4.123105625617661 | +-----------+----------+---------+-----------------+--------------------+ 3 rows in set (0.00 sec) ``` + Linux OS プロンプトを終了します。 ```sql EXIT ``` ```output Bye ``` ### モジュール 1 のまとめ + この演習では、レコードは全部で 3 つしかありません。ソートまたは検索する必要があるベクトルレコードが多数ある場合、効率をどのように確保すればよいでしょうか。次の演習では、この問題を解決できる HNSW インデックスを紹介します。 + ベクトルデータ型を使用して TiDB テーブルを作成する演習をし、ベクトルデータを含むレコードを正常に挿入しました。 + TiDB では、次のベクトル関数がサポートされています: `VEC_L1_DISTANCE`、`VEC_L2_DISTANCE`、`VEC_COSINE_DISTANCE`、`VEC_NEGATIVE_INNER_PRODUCT`、`VEC_DIMS`、`VEC_AS_TEXT`、`VEC_FROM_TEXT`、`VEC_L2_NORM` など。 ## モジュール 2: HNSW インデックスを使用して K 近傍クエリのパフォーマンスを向上させる ![54%](https://lab-static.pingcap.com/quick-demo/TiDBVectorSimilaritySearch_aa_001_20240610.png) ### 参照 + HNSW (Hierarchical Navigable Small World) インデックスは、近似最近傍検索アルゴリズムの一種です。 + HNSWインデックスは、高次元データポイントをグラフ構造で整理します。このグラフ構造では、各ノードがデータポイントを表し、高次元空間で互いに近いノード同士がエッジで結ばれています。 + HNSW インデックスは、このグラフを通じて階層検索を実行することで最近傍クエリを効率的に近似し、高次元空間でも最近傍を高速に取得できるようにします。これは、推奨システム、類似性検索、情報検索などのアプリケーションでよく使用されます。 ### 手順 1. TiDB Cloud Serverless データベースに接続する + 以下の操作はリモート Linux 上で行います。以下のコマンドを使用して TiDB Cloud Serverless に接続します。(MySQL クライアントを終了する場合は `EXIT` を使用します。) ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ -D vs_example \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt ``` ```output mysql: [Warning] Using a password on the command line interface can be insecure. Welcome to the MySQL monitor. Commands end with ; or \g. Your MySQL connection id is 2168969 Server version: 5.7.28-TiDB-Serverless TiDB Server (Apache License 2.0) Community Edition, MySQL 5.7 compatible Copyright (c) 2000, 2024, Oracle and/or its affiliates. Oracle is a registered trademark of Oracle Corporation and/or its affiliates. Other names may be trademarks of their respective owners. Type 'help;' or '\h' for help. Type '\c' to clear the current input statement. mysql> ``` ### 手順 2. ベクトルデータ型のテストテーブルを作成する + サンプル テーブル `vector_table_noidx` にデータ (32768 行) を入力します。 ```sql USE vs_example; DROP SEQUENCE IF EXISTS vs_seq1; CREATE SEQUENCE vs_seq1; DROP TABLE IF EXISTS vector_table_noidx; CREATE TABLE vector_table_noidx ( id BIGINT AUTO_RANDOM PRIMARY KEY, doc TEXT, embedding VECTOR(3)); INSERT INTO vector_table_noidx (doc, embedding) VALUES ('dog', '[-1, -2, -1]'); INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; ANALYZE TABLE vector_table_noidx; INSERT INTO vector_table_noidx (doc, embedding) SELECT NEXTVAL(vs_seq1), VEC_FROM_TEXT(CONCAT('[',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),',',CONVERT(RAND(), CHAR),']')) FROM vector_table_noidx; ANALYZE TABLE vector_table_noidx; ``` ```output ... Query OK, 16384 rows affected (0.06 sec) Records: 16384 Duplicates: 0 Warnings: 0 Query OK, 0 rows affected (0.10 sec) ``` ### 手順 3. K 近傍検索を実行し、実行プランを調べる + 次のクエリを確認してください。これは、コサイン距離を使用した典型的な K 近傍検索を示しています。 + アルゴリズムは、データセット - テーブル `vector_table_noidx` 内の参照ノード `[-1,-1,-1]` に最も近いノードを検索します。 ```sql SELECT doc, embedding FROM vector_table_noidx ORDER BY VEC_COSINE_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +------+------------+ | doc | embedding | +------+------------+ | dog | [-1,-2,-1] | +------+------------+ 1 row in set (0.03 sec) ``` + 実行プランを確認し、クエリで `FullTableScan` が使用されていることを確認します。テーブル `vector_table_noidx` は 32768 行のみで構成されているため、応答時間が非常に速いため、パフォーマンスへの潜在的な影響を考慮することが重要です。特に、影響が大きく、無視できない数十億行を処理する場合はなおさらです。EXPLAIN 出力の下部にある操作に注目してください。推定行数 (`estRows`) は `1.00` よりもはるかに大きい数値です。 ```sql EXPLAIN SELECT doc, embedding FROM vector_table_noidx ORDER BY VEC_COSINE_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +--------------------------------+----------+-----------+--------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------+----------+-----------+--------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Projection_14 | 1.00 | root | | vs_example.vector_table_noidx.doc, vs_example.vector_table_noidx.embedding | | └─TopN_7 | 1.00 | root | | Column#4, offset:0, count:1 | | └─Projection_15 | 1.00 | root | | vs_example.vector_table_noidx.doc, vs_example.vector_table_noidx.embedding, vec_cosine_distance(vs_example.vector_table_noidx.embedding, [-1,-1,-1])->Column#4 | | └─TableReader_13 | 1.00 | root | | data:TopN_12 | | └─TopN_12 | 1.00 | cop[tikv] | | vec_cosine_distance(vs_example.vector_table_noidx.embedding, [-1,-1,-1]), offset:0, count:1 | | └─TableFullScan_11 | 10000.00 | cop[tikv] | table:vector_table_noidx | keep order:false, stats:pseudo | +--------------------------------+----------+-----------+--------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------+ 6 rows in set (0.01 sec) ``` ### 手順 4. ベクトル データ型の HNSW インデックスを持つテーブルを作成する + 列 `embedding` に HNSW インデックスが定義された新しいテーブル `vector_table_with_index_cos` を作成し、テーブル `vector_table_noidx` の同じデータを入力します。HNSW インデックスは、近似精度 (ほとんどの場合 98% 以上) で優れたパフォーマンスを提供します。HNSW インデックスを作成するための `COMMENT "HNSW()"` 構文は、将来変更される可能性があることに注意してください。 + `cosine distance` HNSW インデックスを持つテーブルを作成します。 ```sql DROP TABLE IF EXISTS vector_table_with_index_cos; CREATE TABLE vector_table_with_index_cos ( id BIGINT AUTO_RANDOM PRIMARY KEY, doc TEXT, embedding VECTOR(3), VECTOR INDEX embedding_idx ((VEC_COSINE_DISTANCE(embedding))) ); INSERT INTO vector_table_with_index_cos (doc, embedding) SELECT doc, embedding FROM vector_table_noidx; ANALYZE TABLE vector_table_with_index_cos; ``` ```output ... Query OK, 32768 rows affected (0.03 sec) Records: 32768 Duplicates: 0 Warnings: 0 Query OK, 0 rows affected (0.08 sec) ``` + HNSW インデックスは TiFlash コピーに依存しているため、TiFlash コピーが存在するかどうかを確認してください。示されているのと同じ出力が表示され、`AVAILABLE` と `PROGRESS` が両方とも 1 である場合、コピーは所定の位置にあることに注意してください。値が 1 未満の場合は、次のコマンドを再度実行して進行状況を確認できます。コピー速度は非常に速くなります。 ```sql SELECT * FROM information_schema.tiflash_replica WHERE TABLE_SCHEMA = 'vs_example'; ``` ```output +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ | TABLE_SCHEMA | TABLE_NAME | TABLE_ID | REPLICA_COUNT | LOCATION_LABELS | AVAILABLE | PROGRESS | +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ | vs_example | vector_table_with_index_cos | 183 | 2 | | 1 | 1 | +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ ``` ### 手順 5. インデックス付きテーブルに対して K 近傍クエリを実行する + テーブル `vector_table_with_index_cos` に対して同じ k 近傍検索を実行します。 ```sql SELECT doc, embedding FROM vector_table_with_index_cos ORDER BY VEC_COSINE_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +------+------------+ | doc | embedding | +------+------------+ | dog | [-1,-2,-1] | +------+------------+ 1 row in set (0.03 sec) ``` ### 手順 6. 実行プランを確認する + 実行プランを確認し、近似最近傍インデックス (出力の `access object` 列の `index:embedding_idx(embedding)` オブジェクトと `operator info` 列の `annIndex:COSINE` 演算子) がクエリで使用されていることを確認します。大規模に運用されている本番環境では、このクエリを処理するときにテーブル `vector_table_with_index_cos` が `vector_table_noidx` よりもパフォーマンスが優れていることが期待されます。explain 出力の下部にある操作に注目してください。推定行数 (`estRows`) は `1.00` のみであり、これはステップ 2 よりも大幅に効率的です。インデックスの最適化に加えて、出力の `task` 列の `mpp[tiflash]` タスクも見つかります。これは、TiDB によって実装された別の列指向ストレージ クエリの最適化です。 ```sql EXPLAIN SELECT doc, embedding FROM vector_table_with_index_cos ORDER BY VEC_COSINE_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +--------------------------------------+---------+--------------+-------------------------------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------------+---------+--------------+-------------------------------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Projection_28 | 1.00 | root | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding | | └─TopN_10 | 1.00 | root | | Column#5, offset:0, count:1 | | └─Projection_29 | 1.00 | root | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding, vec_cosine_distance(vs_example.vector_table_with_index_cos.embedding, [-1,-1,-1])->Column#5 | | └─TableReader_25 | 1.00 | root | | MppVersion: 1, data:ExchangeSender_24 | | └─ExchangeSender_24 | 1.00 | mpp[tiflash] | | ExchangeType: PassThrough | | └─Projection_26 | 1.00 | mpp[tiflash] | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding | | └─TopN_23 | 1.00 | mpp[tiflash] | | Column#4, offset:0, count:1 | | └─Projection_27 | 1.00 | mpp[tiflash] | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding, vec_cosine_distance(vs_example.vector_table_with_index_cos.embedding, [-1,-1,-1])->Column#4 | | └─TableFullScan_22 | 1.00 | mpp[tiflash] | table:vector_table_with_index_cos, index:embedding_idx(embedding) | keep order:false, stats:pseudo, annIndex:COSINE(embedding..[-1,-1,-1], limit:1) | +--------------------------------------+---------+--------------+-------------------------------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ 9 rows in set (0.01 sec) ``` ### 手順 7. L2 距離インデックスを試す + `コサイン距離` に加えて、`L2 距離` インデックスもサポートされています。 + 今回は検索アルゴリズムが `L2 距離` に変更されたため、実行プランを確認して、テーブル `vector_table_with_index_cos` に対するクエリで近似最近傍インデックス (`annIndex`) が使用されていないこと、代わりに `TableFullScan` が使用されていることを確認します。 ```sql EXPLAIN SELECT doc, embedding FROM vector_table_with_index_cos ORDER BY VEC_L2_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +--------------------------------------+----------+--------------+-----------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------------+----------+--------------+-----------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Projection_28 | 1.00 | root | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding | | └─TopN_9 | 1.00 | root | | Column#5, offset:0, count:1 | | └─Projection_29 | 1.00 | root | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding, vec_l2_distance(vs_example.vector_table_with_index_cos.embedding, [-1,-1,-1])->Column#5 | | └─TableReader_20 | 1.00 | root | | MppVersion: 1, data:ExchangeSender_19 | | └─ExchangeSender_19 | 1.00 | mpp[tiflash] | | ExchangeType: PassThrough | | └─Projection_26 | 1.00 | mpp[tiflash] | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding | | └─TopN_18 | 1.00 | mpp[tiflash] | | Column#4, offset:0, count:1 | | └─Projection_27 | 10000.00 | mpp[tiflash] | | vs_example.vector_table_with_index_cos.doc, vs_example.vector_table_with_index_cos.embedding, vec_l2_distance(vs_example.vector_table_with_index_cos.embedding, [-1,-1,-1])->Column#4 | | └─TableFullScan_16 | 10000.00 | mpp[tiflash] | table:vector_table_with_index_cos | keep order:false, stats:pseudo | +--------------------------------------+----------+--------------+-----------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ 9 rows in set (0.88 sec) ``` + `L2 距離` HNSW インデックスを持つテーブルを作成し、テーブル `vector_table_with_index_cos` と同じデータを入力します。 ```sql DROP TABLE IF EXISTS vector_table_with_index_l2; CREATE TABLE vector_table_with_index_l2 ( id BIGINT AUTO_RANDOM PRIMARY KEY, doc TEXT, embedding VECTOR(3), VECTOR INDEX embedding_idx ((VEC_L2_DISTANCE(embedding))) ); INSERT INTO vector_table_with_index_l2 (doc, embedding) SELECT doc, embedding FROM vector_table_with_index_cos; ANALYZE TABLE vector_table_with_index_l2; ``` ```output Query OK, 32768 rows affected (0.62 sec) Records: 32768 Duplicates: 0 Warnings: 0 Query OK, 0 rows affected (0.09 sec) ``` + HNSW インデックスは TiFlash コピーに依存しているため、TiFlash コピーが存在するかどうかを確認してください。今回はテーブル `vector_table_with_index_l2` の `AVAILABLE` と `PROGRESS` に注目してください。 ```sql SELECT * FROM information_schema.tiflash_replica WHERE TABLE_SCHEMA = 'vs_example'; ``` ```output +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ | TABLE_SCHEMA | TABLE_NAME | TABLE_ID | REPLICA_COUNT | LOCATION_LABELS | AVAILABLE | PROGRESS | +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ | vs_example | vector_table_with_index_cos | 183 | 2 | | 1 | 1 | | vs_example | vector_table_with_index_l2 | 186 | 2 | | 1 | 1 | +--------------+-----------------------------+----------+---------------+-----------------+-----------+----------+ 2 rows in set (0.07 sec) ``` + テーブル `vector_table_with_index_l2` に対してクエリを実行し、実行プランをチェックして、EXPLAIN 出力テーブルの下部にある演算子が L2 距離に近似最近傍インデックス (`index:embedding_idx(embedding)`) を使用していることを確認します。 ```sql SELECT doc, embedding FROM vector_table_with_index_l2 ORDER BY VEC_L2_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +------+------------+ | doc | embedding | +------+------------+ | dog | [-1,-2,-1] | +------+------------+ 1 row in set (0.10 sec) ``` ```sql EXPLAIN SELECT doc, embedding FROM vector_table_with_index_l2 ORDER BY VEC_L2_DISTANCE(embedding, '[-1, -1, -1]') LIMIT 1; ``` ```output +--------------------------------------+---------+--------------+------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------------+---------+--------------+------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Projection_28 | 1.00 | root | | vs_example.vector_table_with_index_l2.doc, vs_example.vector_table_with_index_l2.embedding | | └─TopN_10 | 1.00 | root | | Column#5, offset:0, count:1 | | └─Projection_29 | 1.00 | root | | vs_example.vector_table_with_index_l2.doc, vs_example.vector_table_with_index_l2.embedding, vec_l2_distance(vs_example.vector_table_with_index_l2.embedding, [-1,-1,-1])->Column#5 | | └─TableReader_25 | 1.00 | root | | MppVersion: 1, data:ExchangeSender_24 | | └─ExchangeSender_24 | 1.00 | mpp[tiflash] | | ExchangeType: PassThrough | | └─Projection_26 | 1.00 | mpp[tiflash] | | vs_example.vector_table_with_index_l2.doc, vs_example.vector_table_with_index_l2.embedding | | └─TopN_23 | 1.00 | mpp[tiflash] | | Column#4, offset:0, count:1 | | └─Projection_27 | 1.00 | mpp[tiflash] | | vs_example.vector_table_with_index_l2.doc, vs_example.vector_table_with_index_l2.embedding, vec_l2_distance(vs_example.vector_table_with_index_l2.embedding, [-1,-1,-1])->Column#4 | | └─TableFullScan_22 | 1.00 | mpp[tiflash] | table:vector_table_with_index_l2, index:embedding_idx(embedding) | keep order:false, stats:pseudo, annIndex:L2(embedding..[-1,-1,-1], limit:1) | +--------------------------------------+---------+--------------+------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ 9 rows in set (0.01 sec) ``` + MySQL クライアントを終了し、リモート Linux OS プロンプトに戻ります。 ```sql EXIT ``` ```output Bye ``` ### モジュール 2 のまとめ + この演習では、大規模なデータセットで K 近傍検索のパフォーマンスを確保するために、TiDB テーブルに HNSW インデックスを作成する方法を演習しました。 + TiDB では、次の HNSW がサポートされています: `Cosine Distance`、`Square Distance (L2)`。 + TiDB をベクトル ストアとして使用すると、次のような利点がすぐに得られます: + 効率的な近似最近傍検索。 + 計画的および計画外のダウンタイムはゼロです。(TiDB のゼロ ダウンタイム機能の詳細については、他のハンズオン 演習を参照してください。) + リレーショナル データとベクトル データの両方に対応する統合ストレージで、非常に大きなボリュームまで拡張可能です。 + 列指向ストレージ TiFlash の詳細については、他のハンズオン 演習「TiDB Cloud Serverless 入門」を確認するか、自習型ラーニング「TiDB の紹介」(https://pingcap.co.jp/education/course-introduction-to-tidb/) に登録してください。 + 次の演習では、TiDB のベクトル検索機能を適用して実際の問題を解決します。 ## モジュール 3: TiDB を使用して RAG アプリケーションのセマンティック検索をサポートする + 最初の 2 つのデモで概念とスキルを習得したら、次は実際の例に進みます。以前のデモの 3 次元ベクトルとは異なり、今回は RAG アプリケーションに 1536 次元ベクトルを使用します。このデモでは、RAG アプリケーションはベクトル ストアとして TiDB によってサポートされ、OpenAI が LLM を提供します。 + この練習では、OpenAI API キーが必要です。 ![48%](https://lab-static.pingcap.com/quick-demo/TiDBAsVectorStoreForRAGDetails_aa_001_20240610.png) ### ヒント + RAG はプロンプト強化方法として機能し、プロンプト内に追加のコンテキストを組み込むことで、大規模言語モデル (LLM) がユーザーフレンドリーな応答を生成する機能を強化します。 + 追加のコンテキストでプロンプトを強化するための重要なステップは、セマンティック検索を実行することにあります。これには、TiDB などのベクトル ストアで K 近傍検索を実行することが含まれます。 ### 手順 1. TiDB Cloud Serverless データベースに接続する + 指定された TiDB Cloud Serverless インスタンスに接続し、スキーマ `semantic_search_test` がプロビジョニングされていることを確認します。 ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt ``` ### 手順 2. 既存のスキーマを確認する + データベースにはテーブル `llama_index_rag_test` が 1 つだけあります。 ```sql USE semantic_search_test; SHOW TABLES; ``` ```output +--------------------------------+ | Tables_in_semantic_search_test | +--------------------------------+ | llama_index_rag_test | +--------------------------------+ 1 row in set (0.00 sec) ``` + テーブル スキーマを確認します。列 `embedding` は列 `document` のベクトル埋め込みであり、列 `embedding` には `Cosine Distance` インデックスがあることに注意してください。 ```sql SHOW CREATE TABLE llama_index_rag_test\G ``` ```output *************************** 1. row *************************** Table: llama_index_rag_test Create Table: CREATE TABLE `llama_index_rag_test` ( `id` varchar(36) NOT NULL, `embedding` vector(1536) NOT NULL COMMENT 'hnsw(distance=cosine)', `document` text DEFAULT NULL, `meta` json DEFAULT NULL, `create_time` datetime DEFAULT CURRENT_TIMESTAMP, `update_time` datetime DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP, PRIMARY KEY (`id`) /*T![clustered_index] CLUSTERED */ ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin 1 row in set (0.01 sec) ``` + テーブル `llama_index_rag_test` には、`https://docs.pingcap.com` からクロールされた 99 のドキュメントが含まれており、TiDB バージョン 8.1 までの次のトピックをカバーしています: TiDB の概要、TiDB クイック スタート、TiDB HTAP クイック スタート、TiDB への移行。これらのレコードは、LlamaIndex フレームワークとカスタム構成された OpenAI API キーを使用して作成されました。PingCAP トレーニングおよび認定チームがこの処理を処理し、データをプロビジョニングしましたが、OpenAI アカウントがあれば、後続の手順で自分で実行できます。 ```command SELECT COUNT(DISTINCT document), COUNT(DISTINCT embedding) FROM llama_index_rag_test; ``` ```output +--------------------------+---------------------------+ | COUNT(DISTINCT document) | COUNT(DISTINCT embedding) | +--------------------------+---------------------------+ | 99 | 99 | +--------------------------+---------------------------+ 1 row in set (0.02 sec) ``` + `document` 列の平均長 (3,000 バイト以上) と `embedding` 列の埋め込みベクトルの次元数 (埋め込みごとに 1536 次元) を確認します。 ```sql SELECT AVG(LENGTH(document)) AS avg_document_length, MAX(VEC_DIMS(embedding)) AS max_no_of_dimensions, MIN(VEC_DIMS(embedding)) AS min_no_of_dimensions FROM llama_index_rag_test; ``` ```output +---------------------+----------------------+----------------------+ | avg_document_length | max_no_of_dimensions | min_no_of_dimensions | +---------------------+----------------------+----------------------+ | 3518.3333 | 1536 | 1536 | +---------------------+----------------------+----------------------+ 1 row in set (0.01 sec) ``` + 基本的な理解を得るために、レコードの 1 つを調べます。 ```sql SELECT SUBSTR(document,1,57) AS CONTENT, ' ', embedding AS EMBEDDING FROM llama_index_rag_test ORDER BY create_time LIMIT 1\G ``` ```output *************************** 1. row *************************** CONTENT: Users can elastically scale TiDB to meet the requirements : EMBEDDING: [-0.0027229728,-0.016953643,0.023380065,-0.021632409,-0.009694677,0.020834267,-0.02741206,0.0009624152,-0.029558789,-0.036274195,-0.0010510021,0.023999315,0.0132725565,0.001702933,-0.004558356,0.017504087,0.0037292512,-0.003866862,-0.004056077,-0.022802101,-0.0114767365,0.0022619767,-0.0079951845,0.004035435,-0.021536082,0.008903416,0.047283053,-0.009667155,0.0038634217,-0.000537972,0.03498065,-0.02752215,-0.019237982,0.004045756,-0.010135031,-0.006708524,0.0098254075, . . . -0.0036398042,0.0026696485,0.023063561,-0.022760818,-0.006822053,-0.020820506,-0.047971107,-0.013912447,0.04571429,-0.0026042834,0.0021794103,-0.0511912,0.009082309,-0.026283653,-0.025636882,-0.019361833,-0.020297585,0.0021484478,-0.022636969,-0.024811218,0.011655631,0.00072159636,0.024687368,-0.004241851,-0.018976523,-0.014270235,0.014696828,0.0033301802,0.014724351,-0.015591298,-0.0116212275] 1 row in set (0.01 sec) ``` ### 手順 3. RAG アプリケーションに埋め込まれたベクトルを調べる + 埋め込みベクトル (たとえば、TiDB に関する質問から) がある場合、デモ 2 で示したものと同様の k 最近傍検索を使用してコサイン類似度検索を実行できます。違いは次元にあります。デモ 2 では 3 次元で作業しましたが、今回は 1536 次元で作業しています。 + たとえば、次の 1536 次元ベクトルは、OpenAI の `text-embedding-3-small` モデルを使用して、質問 `MySQL から TiDB に移行するには?` から埋め込まれています。埋め込みの実行方法の詳細については、この [OpenAI ドキュメント](https://platform.openai.com/docs/guides/embeddings) を確認してください。 ```output [-0.0088212425,-0.02732351,0.008679073,-0.008949871,-0.0052805594,-0.043733865,-0.031520877,0.033985138,-0.037207633,-0.020973299,0.011414132,0.050287172,-0.014704327,-0.003625307,0.000286665,0.078910515,-0.0034797532,-0.0055479724,-0.008922791,0.052318156,0.0006689555,-0.012395775,0.0033528167,-0.015462561,0.0058323103,-0.04015933,-0.051262043,0.017371686,0.03133132,-0.040457208,0.028298382,-0.03625984,-0.013248788,-0.007196455,-0.028190063,0.0127952015,-0.012761352,0.034635052,-0.010574658,0.011373512,0.023843756,0.02859626,0.0006452607,0.017385226,0.051045407,-0.0156115,-0.040592607,-0.001942975,0.011265193,0.036503557,0.0024168715,-0.033091504,-0.0074063228,0.0336331,-0.017601864,-0.03601612,0.023721898,0.030708484,0.03466213,0.019957807,0.027594307,-0.059033945,0.031033441,0.030789724,-0.00779221,0.012653032,-0.061796084,0.001817731,0.032224953,0.017155048,-0.029300334,0.06222936,-0.053780466,0.010046602,0.039644815,0.038480382,-0.021812772,0.016234335,-0.0078396,-0.01170524,0.038480382,-0.008137478,-0.00388595,-0.037288874,0.042406954,-0.018157,-0.028894138,-0.041946597,0.010012753,-0.035528686,-0.019334972,-0.00009361569,0.0024862634,0.033335224,0.034337174,0.046143964,-0.0285421,0.004691574,-0.028271303,0.004918367,0.04909566,0.009328988,-0.008178097,0.011982808,-0.011244884,0.007392783,-0.008245797,-0.005581822,-0.03111468,0.007379243,0.0364494,-0.042027835,-0.008293186,0.024886329,-0.008211947,0.00781252,0.021474274,-0.0026335097,-0.05995466,-0.05451162,0.0012871363,-0.012801971,-0.00023144759,-0.049826816,-0.030356446,-0.030356446,-0.033362303,0.0016129401,-0.06509982,-0.03514957,0.037343033,0.0064517604,-0.011962498,-0.01167139,-0.020580642,-0.03260407,-0.0116104605,0.022164809,-0.045846086,0.072952956,-0.0010611893,0.020052586,0.038128346,-0.009809654,-0.0052128597,0.0077719,-0.00129983,-0.023830216,-0.0058932398,-0.0065194597,0.029056616,0.011556301,-0.054619938,-0.022340829,-0.012463474,-0.026822533,0.03203539,-0.00027756786,-0.009538856,-0.0009858736,-0.010669438,-0.014826186,-0.040348887,0.011089175,-0.039482336,0.037153475,-0.010703288,-0.026389256,-0.07279048,-0.01165785,-0.011881258,0.022963664,-0.05451162,-0.05605517,-0.025563324,-0.04259651,-0.034580894,0.004691574,-0.045575287,0.044762895,0.033281066,0.08595126,-0.035799485,-0.013553436,-0.014162731,0.033497702,0.0004028119,0.0026216623,-0.021799233,-0.033443544,0.029571133,0.061687764,0.003190338,0.028731659,0.022543926,-0.03639524,0.0035711476,0.009708106,-0.018617356,0.020905599,0.012625953,0.04024057,0.012544713,-0.021636754,-0.08719693,-0.008123937,-0.044979535,0.014582467,-0.009606556,-0.051641162,0.01553026,-0.0015477793,-0.0067902575,0.008245797,0.0070881355,0.010554349,0.01696549,0.05207444,-0.019971346,0.012686882,-0.0027689086,0.020716041,0.051099565,0.010581428,0.03777631,0.024046855,0.022801185,-0.04015933,0.026632976,-0.04403174,0.032252032,0.042081997,0.0023762516,-0.0014640013,-0.0008610527,0.06910763,0.00002308658,0.008083317,-0.0050469963,-0.0049149822,-0.050151773,-0.025265446,0.0018532733,-0.00008182117,0.024994647,-0.001661176,0.013336797,0.02986901,0.028894138,0.051207885,-0.0026436646,0.012145286,-0.0073047737,-0.023180302,-0.01704673,-0.043788023,-0.03739719,0.018238239,0.03544745,0.0061606527,-0.010425719,0.003875795,-0.019714087,-0.015435481,-0.043625545,0.013858084,0.035664085,0.03804711,0.04397758,0.029489893,0.023627117,0.039780214,0.002853533,0.018698595,-0.021826312,0.010994395,0.015340702,0.052561875,-0.07409031,0.0731696,-0.07604005,-0.053617988,0.00033595867,0.003170028,-0.0058018453,-0.013289408,0.005822155,-0.0045595597,-0.013478966,-0.01689779,0.013932553,0.03254991,0.021420116,0.001810961,-0.015191763,0.0068274923,0.0025895052,-0.047633354,0.050937086,0.006143728,0.063041754,0.055432335,0.026199698,-0.00583908,-0.012131747,0.0233157,0.01677593,-0.02479155,-0.0035948423,0.012504093,-0.015449021,-0.02079728,-0.0078396,-0.0036523868,-0.028135903,-0.038074188,0.0012236681,0.035474528,0.050178852,0.008076548,0.017222747,-0.042109076,-0.011116254,-0.011251654,-0.0129306,0.049041502,0.06645381,-0.005273789,-0.019077713,0.008123937,0.018820455,-0.051126648,0.0010730367,0.034418415,0.010127842,-0.04698344,0.003434056,-0.009315448,0.03774923,0.008550444,-0.023627117,-0.04925814,-0.057246678,-0.008022388,-0.007128755,-0.06407078,0.014176271,0.021068078,-0.014081491,0.020336924,-0.017764343,0.009863814,-0.00906496,0.025549782,-0.011353202,-0.0003590188,0.009403458,-0.018468417,-0.015746899,0.042136155,-0.03766799,-0.0015740129,0.0024388738,-0.031547956,0.0030312443,-0.0034865232,0.008855091,-0.03514957,-0.024926947,0.04010517,0.015706278,0.006350211,0.023613578,0.007413093,-0.028894138,-0.011217804,-0.048743624,0.021745073,0.060387935,-0.023613578,-0.036232762,-0.0008081625,-0.022543926,0.016450973,0.01821116,-0.03904906,0.021406576,0.04135084,0.009599786,-0.07214057,0.06878267,0.0011195801,0.017886203,-0.010459569,-0.072628,0.020905599,0.025603943,0.046441842,0.015394862,-0.0044444706,0.056705084,-0.013289408,-0.041946597,-0.044708736,-0.009579476,0.012517634,0.022774104,0.016586373,0.03864286,-0.012504093,0.0018498882,0.002464261,0.03238743,0.05459286,0.013397727,0.015977077,0.037586752,-0.011468291,-0.02062126,-0.02859626,0.017331066,-0.0059507843,0.00010414084,0.021081617,-0.02081082,0.056434285,0.01580106,-0.008773852,0.01691133,-0.08270168,0.007135525,0.037234712,-0.00909204,0.023938535,-0.0009926436,-0.0259289,-0.027133951,0.07159897,-0.00043454603,0.035257887,-0.026565276,-0.022963664,0.011285503,0.008841552,-0.013377417,0.0064348355,0.02740475,-0.023586499,-0.033443544,0.026592355,0.0043090717,0.0023864065,-0.038209587,-0.0066683986,0.031954154,-0.0019615924,-0.0057341456,0.005436268,-0.023654198,0.01559796,0.00228655,0.011881258,-0.0063231313,-0.02471031,0.04124252,-0.041432083,0.0079614585,-0.019037094,0.017791422,0.013289408,-0.042244475,0.05058505,0.012260376,0.044546258,-0.009755495,0.02716103,-0.009396688,-0.02213773,-0.004156748,-0.061579447,0.0019057404,0.0037268563,-0.04419422,0.0135669755,-0.026294477,0.1475307,-0.0010552657,0.02068896,-0.024155173,-0.013045689,0.04284023,0.007521412,-0.024954028,-0.016247874,-0.06233768,0.019037094,-0.00040027316,-0.025495624,0.010134612,-0.0649915,-0.01930789,-0.01171201,0.008469204,0.042000756,0.016803011,-0.0030583239,0.0061809625,-0.0035372977,-0.017222747,0.042325713,0.007650041,0.0070881355,-0.01691133,0.0077109705,-0.024141634,-0.0060624885,-0.048391588,0.009660715,-0.036097363,0.014243971,0.0031344858,0.01683009,-0.013377417,-0.023044903,0.030329367,-0.041052964,-0.008103628,-0.064016625,0.023031363,-0.031791676,0.049962215,-0.0017652638,-0.005466733,0.030546004,-0.009545626,0.017967442,0.021189937,0.0050097615,-0.011163644,0.012510864,0.022841804,0.0033426618,0.045169093,0.017520625,-0.0005221322,-0.021013917,0.011881258,0.0045629446,0.00907173,-0.008002078,-0.032143712,-0.0360432,-0.01047988,0.020025507,0.017439386,-0.027648468,0.008665533,0.0009638713,0.0339039,0.01692487,0.0060117138,-0.026700674,0.034987092,-0.014541848,-0.01039864,0.010073682,-0.008990491,0.0011940496,-0.010567889,-0.05589269,-0.015124064,0.027485989,-0.011001165,0.04576485,-0.0495831,0.010107532,0.015191763,0.017398765,0.04013225,-0.02596952,0.031656276,-0.008049468,-0.02716103,-0.032008313,-0.033741422,-0.015042824,0.039671894,-0.022679325,-0.02075666,-0.025265446,-0.003364664,-0.008090088,0.004901442,0.015949998,0.004414006,0.013072769,0.043679703,0.020120285,0.005036841,-0.00019495335,-0.0045054,-0.034418415,-0.000118685646,0.014216891,0.03999685,0.013790384,0.0029398499,0.020932678,-0.012937371,-0.02213773,-0.007250614,0.0031886455,-0.0033528167,-0.025265446,0.014988665,0.017520625,0.0047152685,-0.030248128,-0.050097615,-0.03241451,-0.06569558,0.007243844,0.058925625,0.014216891,0.0018414258,0.06455822,-0.031981234,-0.04560237,-0.009193589,0.0070407456,0.0022019255,0.004806663,0.02207003,-0.0546741,-0.025631022,0.018102841,0.004955602,-0.023654198,0.029246176,-0.021189937,0.016369734,-0.0051519303,0.015787518,-0.008956641,0.06114617,-0.0031988004,0.019605769,0.01697903,-0.0062892814,-0.007264154,-0.017263368,-0.03533913,0.0040958184,-0.02334278,0.03114176,0.0073318537,-0.046550162,-0.010770987,-0.022936583,-0.0151376035,0.010026293,0.050043456,-0.027012091,-0.020472322,-0.03238743,0.02335632,-0.014243971,-0.035880722,-0.0114005925,0.009802884,-0.022801185,0.0064212955,0.008218717,-0.011150104,-0.012131747,0.008117167,-0.02225959,-0.012185906,-0.025102967,-0.008462435,0.0008733233,-0.0009427152,-0.022354368,0.0038588704,0.01944329,-0.003500063,0.008888941,-0.00090971176,-0.008665533,0.030302286,-0.038372066,-0.018725676,0.05605517,-0.0043361518,0.030383525,-0.062770955,0.030627243,-0.023396939,-0.04262359,0.0031412558,0.02602368,-0.009762265,0.0039773444,-0.024263494,-0.012253606,-0.0021968482,0.020011965,0.01695195,0.020431703,-0.012449934,0.036936834,-0.028244222,0.011955728,-0.025942441,-0.018685056,0.038155425,-0.024954028,-0.0014656937,-0.011502141,-0.038236666,0.043733865,0.025563324,-0.004383541,-0.026267398,-0.0077177407,-0.04541281,-0.0141898105,0.009938284,0.039834373,0.037992947,-0.0031361782,0.004674649,-0.03766799,0.01172555,0.017669564,0.03366018,-0.039103217,-0.016302034,-0.025509164,0.027012091,0.0075755715,-0.052507717,0.042027835,0.02316676,-0.0025353455,0.030248128,0.007562032,-0.014000252,-0.042298634,0.0005335565,-0.016342655,0.01039187,0.02724227,-0.0033612791,-0.005327949,-0.012544713,-0.00906496,-0.029083695,0.0469022,0.03466213,0.012382234,-0.0072099944,-0.005070691,0.005981249,0.0036760818,0.012057277,0.03477045,-0.007243844,0.0064179106,0.01550318,-0.012741042,-0.0036997765,-0.072465524,-0.03804711,-0.00519255,-0.0079614585,-0.06542478,-0.028433781,-0.01952453,0.015083443,0.01027678,0.01167816,-0.0065465397,0.000062516236,-0.002117301,-0.019334972,-0.0025031883,-0.01677593,-0.0015367782,-0.03894074,-0.004928522,-0.022476226,-0.04137792,0.0005356721,0.013682065,-0.014216891,0.0011652773,-0.0983538,-0.03918446,0.015164684,-0.022801185,-0.005964324,0.028406702,-0.016383274,-0.0014309977,-0.033226904,-0.009843504,0.007392783,-0.00019474178,-0.041973677,0.040348887,-0.040457208,-0.018157,0.05207444,0.007907299,-0.03753259,-0.010486649,-0.06260848,-0.021393035,-0.008611374,0.044708736,0.0025556553,0.010053373,-0.0046103345,0.013973173,0.035582844,-0.027472448,0.04806663,0.0015807828,-0.021677373,-0.0065194597,-0.011867719,-0.011806789,0.0008369348,-0.019551609,0.042109076,0.014731406,-0.017926821,-0.044654574,0.01557088,0.025441464,-0.04419422,0.0030634014,0.00023504412,0.030925121,0.01565212,0.006340056,0.027567228,0.006725943,0.002520113,-0.032116633,0.0049826815,-0.004820203,-0.00094694644,0.012862901,0.037315954,-0.008916021,0.0040619685,0.017141508,-0.016315574,-0.019984886,-0.006895192,-0.02473739,-0.044627495,-0.014460608,0.011001165,0.0031226384,0.026416337,0.029787771,0.011387052,0.02335632,-0.017358147,-0.021189937,0.0025979674,-0.023464639,0.025319604,0.022652246,0.015963538,-0.0026572046,0.0063095912,0.025157126,0.021108698,-0.025278986,-0.023843756,-0.014934505,-0.009261289,-0.029029537,0.016356194,0.0009020956,0.014541848,0.023884377,0.0009334066,0.0049962215,-0.023762517,0.0053990334,-0.020364003,-0.010182002,-0.0004612027,-0.008211947,0.02984193,-0.016803011,-0.011102715,-0.0079614585,0.019118333,-0.02466969,0.016112477,0.021284716,0.021501355,-0.006979816,0.031223,0.014880345,-0.025590403,0.022354368,-0.009362838,-0.019930726,-0.021731533,-0.009342528,0.005625827,-0.016031237,-0.0011864334,0.0019277426,-0.0040958184,0.00645853,-0.012294225,-0.005463348,0.006624394,-0.0124093145,0.035718244,0.05042257,-0.0007078827,-0.03341646,0.068457715,-0.032143712,0.025048807,0.0126462625,-0.04273191,0.020011965,0.0077109705,-0.02323446,-0.0011627385,-0.03249575,0.006377291,0.019808868,-0.0017347991,0.027662007,-0.021081617,-0.011387052,-0.011475062,0.0012177444,0.013946093,-0.018833995,-0.03918446,0.01944329,-0.052561875,-0.03793879,-0.0036049972,0.00050901546,0.0059981737,-0.00020669497,-0.024304112,-0.015828138,0.022232508,0.017317526,0.009396688,-0.017127968,0.004254912,-0.022422068,0.030708484,0.008103628,0.012341615,-0.020066125,0.022828264,-0.0031514107,0.036476478,-0.030491846,0.040565528,-0.020106746,0.055730212,0.027810946,0.006658244,-0.0042582974,0.0051146955,0.009552397,0.03533913,-0.035501607,0.0025218055,-0.0014233815,0.033064425,-0.0060218684,0.013194628,-0.005490428,0.012206216,0.0076094214,0.013330027,-0.02976069,-0.024371812,0.014731406,0.06807859,0.02062126,0.011495371,-0.0060015586,0.0044580107,0.045575287,-0.015313622,-0.030058568,0.04812079,-0.00010831917,-0.019768247,-0.013722684,0.00077473593,0.014406449,-0.010804837,-0.006208042,-0.048743624,0.007853139,-0.027012091,-0.04067385,0.026646515,-0.03406638,-0.009288369,0.007954689,-0.0054091886,0.009349298,0.031656276,-0.00012630183,-0.016721772,-0.001428459,-0.009796115,-0.0025048808,0.020025507,0.025265446,-0.010818377,0.013573745,-0.0018685056,-0.014799106,-0.010154922,0.0017119505,0.013174319,0.022503307,-0.009626866,0.0028873829,0.003889335,0.00083566544,0.018305939,0.007914069,0.022543926,0.021636754,-0.026010139,-0.0073047737,-0.019998426,-0.012727502,-0.015205303,0.021785691,0.023044903,0.008096858,-0.007886989,0.0094440775,0.026308017,0.014555388,0.022300208,0.0070881355,-0.031656276,-0.018915234,0.00095964014,0.02343756,0.021230556,0.009647176,0.062554315,0.033091504,0.022449147,0.030762643,0.024344733,0.057625797,0.013986712,-0.01292383,0.008807702,0.0022560852,-0.009721645,-0.0027435215,0.014054412,-0.0180893,-0.021555513,-0.022868885,-0.0060421783,0.0032817323,0.021081617,0.00051832414,0.008442125,-0.0052636345,0.036070284,0.021433655,-0.027012091,-0.024480132,-0.0059237042,-0.0012499016,0.02204295,-0.034526736,0.032739468,0.043463066,0.008448895,-0.0038317905,0.0070813657,0.022598086,0.0067360983,0.01031063,0.010588199,-0.0010561119,0.03929278,0.0015333932,-0.01948391,0.029679451,0.007264154,-0.0029838546,0.007521412,0.018725676,-0.022246048,0.025021728,-0.014406449,0.017452925,-0.010506959,-0.012097897,-0.014920965,0.009796115,-0.016423894,-0.014406449,-0.0006689555,-0.009667486,0.014839726,0.001143275,0.010324171,0.026362177,0.034987092,0.02458845,-0.005327949,0.012328075,0.01033771,-0.00903788,-0.020418163,-0.0052263997,0.008293186,0.01167139,-0.011197493,0.007501102,-0.023654198,0.0002968199,-0.0013429885,0.0005703681,0.10311984,-0.007534952,-0.004813433,0.027865106,0.008313496,-0.00054201897,0.027120411,-0.03986145,-0.002091914,-0.005405803,-0.03625984,0.018536117,0.021948172,0.016992569,0.011177184,0.026632976,-0.008205177,-0.008828012,-0.0050943857,0.0029144625,0.0012499016,0.025184207,-0.018793376,0.012849361,0.024317652,-0.0019548223,-0.014162731,-0.016261416,0.027580768,0.019186033,-0.0010501882,-0.02878582,0.0046137194,-0.044898294,-0.034391336,0.014636627,0.018861074,0.01683009,0.032766547,-0.03106052,0.0022476227,-0.012659802,0.029923169,0.0038588704,-0.021081617,-0.004133053,0.005290714,0.0013302948,0.005311024,0.010554349,-0.013749764,-0.005273789,-0.007954689,-0.00035499915,-0.032874867,0.010798067,-0.023478178,-0.03999685,0.023329241,0.021812772,-0.025414385,-0.014772026,0.026917312,0.008503054,0.06120033,0.0028586106,-0.029164935,-0.005585207,0.015949998,0.013776843,0.002015752,0.008909252,-0.016748851,0.014149191,0.01429813,0.0129306,-0.008597834,0.0016738696,-0.01432521,0.0062994366,0.024046855,0.017006109,0.03249575,-0.018251779,0.03918446,-0.05345551,0.01158338,0.023694817,-0.0055005825,0.005321179,0.00649915,-0.0065905442,0.0033917439,0.029354494,0.0041905977,0.004705114,-0.0126191825,-0.019876568,0.001073883,0.0040551987,0.018549656,-0.007927609,0.010852226,0.028839977,-0.00086528395,-0.03219787,0.0006232583,0.00052001665,-0.031520877,0.026443416,0.013919013,-0.014975125,-0.0035643775,-0.0047118836,0.0001875487,-0.010182002,-0.02207003,-0.04411298,-0.046252284,0.0129983,-0.009437308,-0.031845834,0.015814599,0.0014174578,-0.008577524,-0.011326123,0.042325713,0.012605643,0.025102967,0.010818377,-0.010242932,-0.014162731,0.023708357,-0.0088821715,0.032143712,-0.032252032,-0.00095710135,-0.0014436914,0.03742427,0.031629197,0.011258423,-0.016383274,-0.017561244,-0.012666573,0.03379558,0.046793878,0.012280686,0.024222873,-0.049528938,-0.0337685,-0.01695195,0.017601864,-0.021731533,-0.017222747,0.04015933,-0.018617356,-0.0014267665,-0.026619434,0.0035203728,0.0019954422,0.02355942,0.0011068864,0.0013700682,0.012829051,-0.012118206,0.0131810885,0.033389382,-0.021365955,-0.018427799,0.011292273,-0.01288998,0.021040998,-0.011894799,0.014758486,0.010798067,0.013390956,0.002836608,0.022638706,-0.024006234,-0.0412696,-0.032901946,-0.03988853,-0.027702626,0.016464513,-0.016518673,0.028487941,-0.014704327,-0.015963538,0.007270924,-0.0004265067,0.04684804,-0.0012245143,0.021122238,0.01041895,-0.01165785,-0.014541848,-0.013878393,-0.035745326,0.0073521635,-0.014731406,0.007000126,0.013458656,-0.0009782575,0.03406638,0.038128346,-0.023708357,0.017520625,-0.034499653,-0.008259336,0.018481957,-0.009559167,0.019050634,-0.0019395901,0.002335632,0.0088821715,0.022963664,-0.008455665,0.036638957,0.0018820455,0.006749638,-0.04316519,-0.036205683,0.003774246,-0.010087222,0.010459569,-0.003347739,-0.018874615,-0.017926821,0.02487279,0.010974086,0.011529221,-0.030627243,0.017452925,-0.001890508,0.02859626,-0.026321556,-0.0051350053,-0.015421941,-0.008949871,-0.011129795,-0.019362051,0.007947919,0.010093993,-0.018779835,-0.008651993,0.05968386,0.019064173,-0.04441086,0.018035142,0.027106872,0.0140408715,-0.014663707,0.005876315,-0.0036930065,-0.0033731265,-0.0039468794,0.015462561,-0.033091504,-0.006993356,0.0042278324,0.03639524,0.026186159,-0.010676208,-0.00015539146,0.022232508,0.030898042,-0.02736413,0.025116507,0.015083443,-0.022218969,-0.01175263,-0.0026233548,-0.016803011,0.024249954,0.012463474,0.013763304,-0.040836327,-0.034851693,-0.016220795,0.002655512,-0.0068241074,0.023261542,-0.0064179106,0.0084353555,0.041161284,0.019835947,0.01423043,0.040782165,0.020851439,-0.020364003,-0.011299043,0.023816677,-0.01807576,0.0035508377,0.0151376035,0.01171878,-0.013946093,0.0018431183,-0.000030279649,-0.017290447,-0.0056833713,0.014758486,0.016071856,-0.008814472,0.004437701,-0.0035339128,-0.008056238,0.044058822,0.008550444,-0.00389949,0.028948298,-0.016694691,-0.031277157,0.0042312173,-0.0050266865,0.0005856005,0.010892846,-0.0019683624,0.0022053106,-0.019714087,0.031737514,-0.014975125,0.014352289,0.010513729,0.002679207,0.0011737397,-0.0017432616,0.0015054671,0.023681277,0.03406638,0.025251906,0.010730367,-0.019998426,-0.013099849,0.024899868,0.008536904,0.009484697,0.019334972,0.0233834,-0.019863026,0.030681403,-0.008252567,-0.015096984,0.014122111,0.047200076,0.03382266,0.045006614,-0.018576737,-0.027580768,0.004786353,-0.022882424,-0.0091191195,-0.0336331,-0.023518799,-0.01041895,0.004502015,0.029246176,0.015029284,0.014216891,-0.016450973,-0.024304112,-0.011732319,0.024859248,0.0010975777,0.01292383,0.017588325,-0.018630896,-0.010899616,0.008699383,0.040809244,-0.022747025,-0.020120285,0.007683891,0.011556301,-0.0073724734,0.0012168981,-0.02079728,-0.0064280657,-0.0075552617,-0.0046949587,0.037315954,0.010425719,0.022394987,-0.0020360618,0.0103174,-0.024994647,-0.036097363,-0.008570754,0.028135903] ``` ### 手順 4. 埋め込まれたベクトルに基づいてセマンティック類似性検索を実行する + セマンティック類似性検索を実行して、質問 `How to migrate to TiDB from MySQL?` に最も近い 3 つの `document` を取得します。取得したドキュメントを確認して、「Migration to TiDB from MySQL」というトピックとの関連性を確認します。それらを LLM の追加のコンテキスト プロンプトとして提供することで、質問に対してより正確な回答を提供する能力を高めることができます。それらは `TiDB Data Migration` セクションからのものであることがわかります。簡潔さを確保するために、出力は切り捨てられています。 ```sql PAGER less -S SELECT document FROM llama_index_rag_test ORDER BY VEC_COSINE_DISTANCE(embedding, '[-0.0088212425,-0.02732351,0.008679073,-0.008949871,-0.0052805594,-0.043733865,-0.031520877,0.033985138,-0.037207633,-0.020973299,0.011414132,0.050287172,-0.014704327,-0.003625307,0.000286665,0.078910515,-0.0034797532,-0.0055479724,-0.008922791,0.052318156,0.0006689555,-0.012395775,0.0033528167,-0.015462561,0.0058323103,-0.04015933,-0.051262043,0.017371686,0.03133132,-0.040457208,0.028298382,-0.03625984,-0.013248788,-0.007196455,-0.028190063,0.0127952015,-0.012761352,0.034635052,-0.010574658,0.011373512,0.023843756,0.02859626,0.0006452607,0.017385226,0.051045407,-0.0156115,-0.040592607,-0.001942975,0.011265193,0.036503557,0.0024168715,-0.033091504,-0.0074063228,0.0336331,-0.017601864,-0.03601612,0.023721898,0.030708484,0.03466213,0.019957807,0.027594307,-0.059033945,0.031033441,0.030789724,-0.00779221,0.012653032,-0.061796084,0.001817731,0.032224953,0.017155048,-0.029300334,0.06222936,-0.053780466,0.010046602,0.039644815,0.038480382,-0.021812772,0.016234335,-0.0078396,-0.01170524,0.038480382,-0.008137478,-0.00388595,-0.037288874,0.042406954,-0.018157,-0.028894138,-0.041946597,0.010012753,-0.035528686,-0.019334972,-0.00009361569,0.0024862634,0.033335224,0.034337174,0.046143964,-0.0285421,0.004691574,-0.028271303,0.004918367,0.04909566,0.009328988,-0.008178097,0.011982808,-0.011244884,0.007392783,-0.008245797,-0.005581822,-0.03111468,0.007379243,0.0364494,-0.042027835,-0.008293186,0.024886329,-0.008211947,0.00781252,0.021474274,-0.0026335097,-0.05995466,-0.05451162,0.0012871363,-0.012801971,-0.00023144759,-0.049826816,-0.030356446,-0.030356446,-0.033362303,0.0016129401,-0.06509982,-0.03514957,0.037343033,0.0064517604,-0.011962498,-0.01167139,-0.020580642,-0.03260407,-0.0116104605,0.022164809,-0.045846086,0.072952956,-0.0010611893,0.020052586,0.038128346,-0.009809654,-0.0052128597,0.0077719,-0.00129983,-0.023830216,-0.0058932398,-0.0065194597,0.029056616,0.011556301,-0.054619938,-0.022340829,-0.012463474,-0.026822533,0.03203539,-0.00027756786,-0.009538856,-0.0009858736,-0.010669438,-0.014826186,-0.040348887,0.011089175,-0.039482336,0.037153475,-0.010703288,-0.026389256,-0.07279048,-0.01165785,-0.011881258,0.022963664,-0.05451162,-0.05605517,-0.025563324,-0.04259651,-0.034580894,0.004691574,-0.045575287,0.044762895,0.033281066,0.08595126,-0.035799485,-0.013553436,-0.014162731,0.033497702,0.0004028119,0.0026216623,-0.021799233,-0.033443544,0.029571133,0.061687764,0.003190338,0.028731659,0.022543926,-0.03639524,0.0035711476,0.009708106,-0.018617356,0.020905599,0.012625953,0.04024057,0.012544713,-0.021636754,-0.08719693,-0.008123937,-0.044979535,0.014582467,-0.009606556,-0.051641162,0.01553026,-0.0015477793,-0.0067902575,0.008245797,0.0070881355,0.010554349,0.01696549,0.05207444,-0.019971346,0.012686882,-0.0027689086,0.020716041,0.051099565,0.010581428,0.03777631,0.024046855,0.022801185,-0.04015933,0.026632976,-0.04403174,0.032252032,0.042081997,0.0023762516,-0.0014640013,-0.0008610527,0.06910763,0.00002308658,0.008083317,-0.0050469963,-0.0049149822,-0.050151773,-0.025265446,0.0018532733,-0.00008182117,0.024994647,-0.001661176,0.013336797,0.02986901,0.028894138,0.051207885,-0.0026436646,0.012145286,-0.0073047737,-0.023180302,-0.01704673,-0.043788023,-0.03739719,0.018238239,0.03544745,0.0061606527,-0.010425719,0.003875795,-0.019714087,-0.015435481,-0.043625545,0.013858084,0.035664085,0.03804711,0.04397758,0.029489893,0.023627117,0.039780214,0.002853533,0.018698595,-0.021826312,0.010994395,0.015340702,0.052561875,-0.07409031,0.0731696,-0.07604005,-0.053617988,0.00033595867,0.003170028,-0.0058018453,-0.013289408,0.005822155,-0.0045595597,-0.013478966,-0.01689779,0.013932553,0.03254991,0.021420116,0.001810961,-0.015191763,0.0068274923,0.0025895052,-0.047633354,0.050937086,0.006143728,0.063041754,0.055432335,0.026199698,-0.00583908,-0.012131747,0.0233157,0.01677593,-0.02479155,-0.0035948423,0.012504093,-0.015449021,-0.02079728,-0.0078396,-0.0036523868,-0.028135903,-0.038074188,0.0012236681,0.035474528,0.050178852,0.008076548,0.017222747,-0.042109076,-0.011116254,-0.011251654,-0.0129306,0.049041502,0.06645381,-0.005273789,-0.019077713,0.008123937,0.018820455,-0.051126648,0.0010730367,0.034418415,0.010127842,-0.04698344,0.003434056,-0.009315448,0.03774923,0.008550444,-0.023627117,-0.04925814,-0.057246678,-0.008022388,-0.007128755,-0.06407078,0.014176271,0.021068078,-0.014081491,0.020336924,-0.017764343,0.009863814,-0.00906496,0.025549782,-0.011353202,-0.0003590188,0.009403458,-0.018468417,-0.015746899,0.042136155,-0.03766799,-0.0015740129,0.0024388738,-0.031547956,0.0030312443,-0.0034865232,0.008855091,-0.03514957,-0.024926947,0.04010517,0.015706278,0.006350211,0.023613578,0.007413093,-0.028894138,-0.011217804,-0.048743624,0.021745073,0.060387935,-0.023613578,-0.036232762,-0.0008081625,-0.022543926,0.016450973,0.01821116,-0.03904906,0.021406576,0.04135084,0.009599786,-0.07214057,0.06878267,0.0011195801,0.017886203,-0.010459569,-0.072628,0.020905599,0.025603943,0.046441842,0.015394862,-0.0044444706,0.056705084,-0.013289408,-0.041946597,-0.044708736,-0.009579476,0.012517634,0.022774104,0.016586373,0.03864286,-0.012504093,0.0018498882,0.002464261,0.03238743,0.05459286,0.013397727,0.015977077,0.037586752,-0.011468291,-0.02062126,-0.02859626,0.017331066,-0.0059507843,0.00010414084,0.021081617,-0.02081082,0.056434285,0.01580106,-0.008773852,0.01691133,-0.08270168,0.007135525,0.037234712,-0.00909204,0.023938535,-0.0009926436,-0.0259289,-0.027133951,0.07159897,-0.00043454603,0.035257887,-0.026565276,-0.022963664,0.011285503,0.008841552,-0.013377417,0.0064348355,0.02740475,-0.023586499,-0.033443544,0.026592355,0.0043090717,0.0023864065,-0.038209587,-0.0066683986,0.031954154,-0.0019615924,-0.0057341456,0.005436268,-0.023654198,0.01559796,0.00228655,0.011881258,-0.0063231313,-0.02471031,0.04124252,-0.041432083,0.0079614585,-0.019037094,0.017791422,0.013289408,-0.042244475,0.05058505,0.012260376,0.044546258,-0.009755495,0.02716103,-0.009396688,-0.02213773,-0.004156748,-0.061579447,0.0019057404,0.0037268563,-0.04419422,0.0135669755,-0.026294477,0.1475307,-0.0010552657,0.02068896,-0.024155173,-0.013045689,0.04284023,0.007521412,-0.024954028,-0.016247874,-0.06233768,0.019037094,-0.00040027316,-0.025495624,0.010134612,-0.0649915,-0.01930789,-0.01171201,0.008469204,0.042000756,0.016803011,-0.0030583239,0.0061809625,-0.0035372977,-0.017222747,0.042325713,0.007650041,0.0070881355,-0.01691133,0.0077109705,-0.024141634,-0.0060624885,-0.048391588,0.009660715,-0.036097363,0.014243971,0.0031344858,0.01683009,-0.013377417,-0.023044903,0.030329367,-0.041052964,-0.008103628,-0.064016625,0.023031363,-0.031791676,0.049962215,-0.0017652638,-0.005466733,0.030546004,-0.009545626,0.017967442,0.021189937,0.0050097615,-0.011163644,0.012510864,0.022841804,0.0033426618,0.045169093,0.017520625,-0.0005221322,-0.021013917,0.011881258,0.0045629446,0.00907173,-0.008002078,-0.032143712,-0.0360432,-0.01047988,0.020025507,0.017439386,-0.027648468,0.008665533,0.0009638713,0.0339039,0.01692487,0.0060117138,-0.026700674,0.034987092,-0.014541848,-0.01039864,0.010073682,-0.008990491,0.0011940496,-0.010567889,-0.05589269,-0.015124064,0.027485989,-0.011001165,0.04576485,-0.0495831,0.010107532,0.015191763,0.017398765,0.04013225,-0.02596952,0.031656276,-0.008049468,-0.02716103,-0.032008313,-0.033741422,-0.015042824,0.039671894,-0.022679325,-0.02075666,-0.025265446,-0.003364664,-0.008090088,0.004901442,0.015949998,0.004414006,0.013072769,0.043679703,0.020120285,0.005036841,-0.00019495335,-0.0045054,-0.034418415,-0.000118685646,0.014216891,0.03999685,0.013790384,0.0029398499,0.020932678,-0.012937371,-0.02213773,-0.007250614,0.0031886455,-0.0033528167,-0.025265446,0.014988665,0.017520625,0.0047152685,-0.030248128,-0.050097615,-0.03241451,-0.06569558,0.007243844,0.058925625,0.014216891,0.0018414258,0.06455822,-0.031981234,-0.04560237,-0.009193589,0.0070407456,0.0022019255,0.004806663,0.02207003,-0.0546741,-0.025631022,0.018102841,0.004955602,-0.023654198,0.029246176,-0.021189937,0.016369734,-0.0051519303,0.015787518,-0.008956641,0.06114617,-0.0031988004,0.019605769,0.01697903,-0.0062892814,-0.007264154,-0.017263368,-0.03533913,0.0040958184,-0.02334278,0.03114176,0.0073318537,-0.046550162,-0.010770987,-0.022936583,-0.0151376035,0.010026293,0.050043456,-0.027012091,-0.020472322,-0.03238743,0.02335632,-0.014243971,-0.035880722,-0.0114005925,0.009802884,-0.022801185,0.0064212955,0.008218717,-0.011150104,-0.012131747,0.008117167,-0.02225959,-0.012185906,-0.025102967,-0.008462435,0.0008733233,-0.0009427152,-0.022354368,0.0038588704,0.01944329,-0.003500063,0.008888941,-0.00090971176,-0.008665533,0.030302286,-0.038372066,-0.018725676,0.05605517,-0.0043361518,0.030383525,-0.062770955,0.030627243,-0.023396939,-0.04262359,0.0031412558,0.02602368,-0.009762265,0.0039773444,-0.024263494,-0.012253606,-0.0021968482,0.020011965,0.01695195,0.020431703,-0.012449934,0.036936834,-0.028244222,0.011955728,-0.025942441,-0.018685056,0.038155425,-0.024954028,-0.0014656937,-0.011502141,-0.038236666,0.043733865,0.025563324,-0.004383541,-0.026267398,-0.0077177407,-0.04541281,-0.0141898105,0.009938284,0.039834373,0.037992947,-0.0031361782,0.004674649,-0.03766799,0.01172555,0.017669564,0.03366018,-0.039103217,-0.016302034,-0.025509164,0.027012091,0.0075755715,-0.052507717,0.042027835,0.02316676,-0.0025353455,0.030248128,0.007562032,-0.014000252,-0.042298634,0.0005335565,-0.016342655,0.01039187,0.02724227,-0.0033612791,-0.005327949,-0.012544713,-0.00906496,-0.029083695,0.0469022,0.03466213,0.012382234,-0.0072099944,-0.005070691,0.005981249,0.0036760818,0.012057277,0.03477045,-0.007243844,0.0064179106,0.01550318,-0.012741042,-0.0036997765,-0.072465524,-0.03804711,-0.00519255,-0.0079614585,-0.06542478,-0.028433781,-0.01952453,0.015083443,0.01027678,0.01167816,-0.0065465397,0.000062516236,-0.002117301,-0.019334972,-0.0025031883,-0.01677593,-0.0015367782,-0.03894074,-0.004928522,-0.022476226,-0.04137792,0.0005356721,0.013682065,-0.014216891,0.0011652773,-0.0983538,-0.03918446,0.015164684,-0.022801185,-0.005964324,0.028406702,-0.016383274,-0.0014309977,-0.033226904,-0.009843504,0.007392783,-0.00019474178,-0.041973677,0.040348887,-0.040457208,-0.018157,0.05207444,0.007907299,-0.03753259,-0.010486649,-0.06260848,-0.021393035,-0.008611374,0.044708736,0.0025556553,0.010053373,-0.0046103345,0.013973173,0.035582844,-0.027472448,0.04806663,0.0015807828,-0.021677373,-0.0065194597,-0.011867719,-0.011806789,0.0008369348,-0.019551609,0.042109076,0.014731406,-0.017926821,-0.044654574,0.01557088,0.025441464,-0.04419422,0.0030634014,0.00023504412,0.030925121,0.01565212,0.006340056,0.027567228,0.006725943,0.002520113,-0.032116633,0.0049826815,-0.004820203,-0.00094694644,0.012862901,0.037315954,-0.008916021,0.0040619685,0.017141508,-0.016315574,-0.019984886,-0.006895192,-0.02473739,-0.044627495,-0.014460608,0.011001165,0.0031226384,0.026416337,0.029787771,0.011387052,0.02335632,-0.017358147,-0.021189937,0.0025979674,-0.023464639,0.025319604,0.022652246,0.015963538,-0.0026572046,0.0063095912,0.025157126,0.021108698,-0.025278986,-0.023843756,-0.014934505,-0.009261289,-0.029029537,0.016356194,0.0009020956,0.014541848,0.023884377,0.0009334066,0.0049962215,-0.023762517,0.0053990334,-0.020364003,-0.010182002,-0.0004612027,-0.008211947,0.02984193,-0.016803011,-0.011102715,-0.0079614585,0.019118333,-0.02466969,0.016112477,0.021284716,0.021501355,-0.006979816,0.031223,0.014880345,-0.025590403,0.022354368,-0.009362838,-0.019930726,-0.021731533,-0.009342528,0.005625827,-0.016031237,-0.0011864334,0.0019277426,-0.0040958184,0.00645853,-0.012294225,-0.005463348,0.006624394,-0.0124093145,0.035718244,0.05042257,-0.0007078827,-0.03341646,0.068457715,-0.032143712,0.025048807,0.0126462625,-0.04273191,0.020011965,0.0077109705,-0.02323446,-0.0011627385,-0.03249575,0.006377291,0.019808868,-0.0017347991,0.027662007,-0.021081617,-0.011387052,-0.011475062,0.0012177444,0.013946093,-0.018833995,-0.03918446,0.01944329,-0.052561875,-0.03793879,-0.0036049972,0.00050901546,0.0059981737,-0.00020669497,-0.024304112,-0.015828138,0.022232508,0.017317526,0.009396688,-0.017127968,0.004254912,-0.022422068,0.030708484,0.008103628,0.012341615,-0.020066125,0.022828264,-0.0031514107,0.036476478,-0.030491846,0.040565528,-0.020106746,0.055730212,0.027810946,0.006658244,-0.0042582974,0.0051146955,0.009552397,0.03533913,-0.035501607,0.0025218055,-0.0014233815,0.033064425,-0.0060218684,0.013194628,-0.005490428,0.012206216,0.0076094214,0.013330027,-0.02976069,-0.024371812,0.014731406,0.06807859,0.02062126,0.011495371,-0.0060015586,0.0044580107,0.045575287,-0.015313622,-0.030058568,0.04812079,-0.00010831917,-0.019768247,-0.013722684,0.00077473593,0.014406449,-0.010804837,-0.006208042,-0.048743624,0.007853139,-0.027012091,-0.04067385,0.026646515,-0.03406638,-0.009288369,0.007954689,-0.0054091886,0.009349298,0.031656276,-0.00012630183,-0.016721772,-0.001428459,-0.009796115,-0.0025048808,0.020025507,0.025265446,-0.010818377,0.013573745,-0.0018685056,-0.014799106,-0.010154922,0.0017119505,0.013174319,0.022503307,-0.009626866,0.0028873829,0.003889335,0.00083566544,0.018305939,0.007914069,0.022543926,0.021636754,-0.026010139,-0.0073047737,-0.019998426,-0.012727502,-0.015205303,0.021785691,0.023044903,0.008096858,-0.007886989,0.0094440775,0.026308017,0.014555388,0.022300208,0.0070881355,-0.031656276,-0.018915234,0.00095964014,0.02343756,0.021230556,0.009647176,0.062554315,0.033091504,0.022449147,0.030762643,0.024344733,0.057625797,0.013986712,-0.01292383,0.008807702,0.0022560852,-0.009721645,-0.0027435215,0.014054412,-0.0180893,-0.021555513,-0.022868885,-0.0060421783,0.0032817323,0.021081617,0.00051832414,0.008442125,-0.0052636345,0.036070284,0.021433655,-0.027012091,-0.024480132,-0.0059237042,-0.0012499016,0.02204295,-0.034526736,0.032739468,0.043463066,0.008448895,-0.0038317905,0.0070813657,0.022598086,0.0067360983,0.01031063,0.010588199,-0.0010561119,0.03929278,0.0015333932,-0.01948391,0.029679451,0.007264154,-0.0029838546,0.007521412,0.018725676,-0.022246048,0.025021728,-0.014406449,0.017452925,-0.010506959,-0.012097897,-0.014920965,0.009796115,-0.016423894,-0.014406449,-0.0006689555,-0.009667486,0.014839726,0.001143275,0.010324171,0.026362177,0.034987092,0.02458845,-0.005327949,0.012328075,0.01033771,-0.00903788,-0.020418163,-0.0052263997,0.008293186,0.01167139,-0.011197493,0.007501102,-0.023654198,0.0002968199,-0.0013429885,0.0005703681,0.10311984,-0.007534952,-0.004813433,0.027865106,0.008313496,-0.00054201897,0.027120411,-0.03986145,-0.002091914,-0.005405803,-0.03625984,0.018536117,0.021948172,0.016992569,0.011177184,0.026632976,-0.008205177,-0.008828012,-0.0050943857,0.0029144625,0.0012499016,0.025184207,-0.018793376,0.012849361,0.024317652,-0.0019548223,-0.014162731,-0.016261416,0.027580768,0.019186033,-0.0010501882,-0.02878582,0.0046137194,-0.044898294,-0.034391336,0.014636627,0.018861074,0.01683009,0.032766547,-0.03106052,0.0022476227,-0.012659802,0.029923169,0.0038588704,-0.021081617,-0.004133053,0.005290714,0.0013302948,0.005311024,0.010554349,-0.013749764,-0.005273789,-0.007954689,-0.00035499915,-0.032874867,0.010798067,-0.023478178,-0.03999685,0.023329241,0.021812772,-0.025414385,-0.014772026,0.026917312,0.008503054,0.06120033,0.0028586106,-0.029164935,-0.005585207,0.015949998,0.013776843,0.002015752,0.008909252,-0.016748851,0.014149191,0.01429813,0.0129306,-0.008597834,0.0016738696,-0.01432521,0.0062994366,0.024046855,0.017006109,0.03249575,-0.018251779,0.03918446,-0.05345551,0.01158338,0.023694817,-0.0055005825,0.005321179,0.00649915,-0.0065905442,0.0033917439,0.029354494,0.0041905977,0.004705114,-0.0126191825,-0.019876568,0.001073883,0.0040551987,0.018549656,-0.007927609,0.010852226,0.028839977,-0.00086528395,-0.03219787,0.0006232583,0.00052001665,-0.031520877,0.026443416,0.013919013,-0.014975125,-0.0035643775,-0.0047118836,0.0001875487,-0.010182002,-0.02207003,-0.04411298,-0.046252284,0.0129983,-0.009437308,-0.031845834,0.015814599,0.0014174578,-0.008577524,-0.011326123,0.042325713,0.012605643,0.025102967,0.010818377,-0.010242932,-0.014162731,0.023708357,-0.0088821715,0.032143712,-0.032252032,-0.00095710135,-0.0014436914,0.03742427,0.031629197,0.011258423,-0.016383274,-0.017561244,-0.012666573,0.03379558,0.046793878,0.012280686,0.024222873,-0.049528938,-0.0337685,-0.01695195,0.017601864,-0.021731533,-0.017222747,0.04015933,-0.018617356,-0.0014267665,-0.026619434,0.0035203728,0.0019954422,0.02355942,0.0011068864,0.0013700682,0.012829051,-0.012118206,0.0131810885,0.033389382,-0.021365955,-0.018427799,0.011292273,-0.01288998,0.021040998,-0.011894799,0.014758486,0.010798067,0.013390956,0.002836608,0.022638706,-0.024006234,-0.0412696,-0.032901946,-0.03988853,-0.027702626,0.016464513,-0.016518673,0.028487941,-0.014704327,-0.015963538,0.007270924,-0.0004265067,0.04684804,-0.0012245143,0.021122238,0.01041895,-0.01165785,-0.014541848,-0.013878393,-0.035745326,0.0073521635,-0.014731406,0.007000126,0.013458656,-0.0009782575,0.03406638,0.038128346,-0.023708357,0.017520625,-0.034499653,-0.008259336,0.018481957,-0.009559167,0.019050634,-0.0019395901,0.002335632,0.0088821715,0.022963664,-0.008455665,0.036638957,0.0018820455,0.006749638,-0.04316519,-0.036205683,0.003774246,-0.010087222,0.010459569,-0.003347739,-0.018874615,-0.017926821,0.02487279,0.010974086,0.011529221,-0.030627243,0.017452925,-0.001890508,0.02859626,-0.026321556,-0.0051350053,-0.015421941,-0.008949871,-0.011129795,-0.019362051,0.007947919,0.010093993,-0.018779835,-0.008651993,0.05968386,0.019064173,-0.04441086,0.018035142,0.027106872,0.0140408715,-0.014663707,0.005876315,-0.0036930065,-0.0033731265,-0.0039468794,0.015462561,-0.033091504,-0.006993356,0.0042278324,0.03639524,0.026186159,-0.010676208,-0.00015539146,0.022232508,0.030898042,-0.02736413,0.025116507,0.015083443,-0.022218969,-0.01175263,-0.0026233548,-0.016803011,0.024249954,0.012463474,0.013763304,-0.040836327,-0.034851693,-0.016220795,0.002655512,-0.0068241074,0.023261542,-0.0064179106,0.0084353555,0.041161284,0.019835947,0.01423043,0.040782165,0.020851439,-0.020364003,-0.011299043,0.023816677,-0.01807576,0.0035508377,0.0151376035,0.01171878,-0.013946093,0.0018431183,-0.000030279649,-0.017290447,-0.0056833713,0.014758486,0.016071856,-0.008814472,0.004437701,-0.0035339128,-0.008056238,0.044058822,0.008550444,-0.00389949,0.028948298,-0.016694691,-0.031277157,0.0042312173,-0.0050266865,0.0005856005,0.010892846,-0.0019683624,0.0022053106,-0.019714087,0.031737514,-0.014975125,0.014352289,0.010513729,0.002679207,0.0011737397,-0.0017432616,0.0015054671,0.023681277,0.03406638,0.025251906,0.010730367,-0.019998426,-0.013099849,0.024899868,0.008536904,0.009484697,0.019334972,0.0233834,-0.019863026,0.030681403,-0.008252567,-0.015096984,0.014122111,0.047200076,0.03382266,0.045006614,-0.018576737,-0.027580768,0.004786353,-0.022882424,-0.0091191195,-0.0336331,-0.023518799,-0.01041895,0.004502015,0.029246176,0.015029284,0.014216891,-0.016450973,-0.024304112,-0.011732319,0.024859248,0.0010975777,0.01292383,0.017588325,-0.018630896,-0.010899616,0.008699383,0.040809244,-0.022747025,-0.020120285,0.007683891,0.011556301,-0.0073724734,0.0012168981,-0.02079728,-0.0064280657,-0.0075552617,-0.0046949587,0.037315954,0.010425719,0.022394987,-0.0020360618,0.0103174,-0.024994647,-0.036097363,-0.008570754,0.028135903]') LIMIT 3\G ``` + 上 キーと 下 キーを押すと、出力をスクロールできます。`dmctl`、`TiDB Data Migration`、`dm-task.yaml` に関連するコンテンツが想定されます。Q キーを押すと、`mysql>` プロンプトに戻ります。 ### 手順 5. TiDB Cloud Serverless 接続を終了する + これで、ベクトル データベースと類似性検索に関する前提知識はすべて理解できました。次のステップでは、すべてをまとめて、ベクトル ストアとして TiDB、LLM として OpenAI Agent を使用するシンプルな Python RAG QA アプリケーションを構築します。リモート Linux の OS プロンプトに戻ります。 ```sql EXIT; ``` ```output mysql> EXIT; Bye ``` ### 手順 6. RAG アプリケーションの環境を準備する + RAG アプリケーションの環境を準備します。この手順から開始して、LlamaIndex クエリ エンジンを機能させるには、リモート Linux の環境変数に独自の OpenAI API KEY を設定する必要があります。 独自の OpenAI API KEY を提供できない場合は、ハンズオン 演習はここで終了します。ただし、引き続き下にスクロールして、提供された出力を調べてください。 + [OpenAI API キー](https://help.openai.com/en/articles/4936850-where-do-i-find-my-openai-api-key)を設定します。リモート Linux で次のコマンドを実行し、`${OPENAI_API_KEY}` を実際の値に置き換えます。 ```bash export OPENAI_API_KEY=${OPENAI_API_KEY} ``` + リモート Linux に必要なライブラリをインストールします。 ```bash pip install --upgrade pip pip install \ "joblib>=1.3.0" \ "click==8.1.7" \ "PyMySQL==1.1.0" \ "SQLAlchemy==2.0.30" \ "tidb-vector==0.0.9" \ "openai==1.27.0" \ "llama-index==0.10.68" \ "llama-index-readers-web==0.1.23" \ "llama-index-vector-stores-tidbvector==0.1.2" \ "httpx==0.27.2" ``` ```output ... Successfully installed PyYAML-6.0.1 aiohttp-3.9.5 aiosignal-1.3.1 annotated-types-0.7.0 anyio-4.4.0 async-timeout-4.0.3 attrs-23.2.0 beautifulsoup4-4.12.3 certifi-2024.6.2 charset-normalizer-3.3.2 chromedriver-autoinstaller-0.6.4 click-8.1.7 cssselect-1.2.0 dataclasses-json-0.6.6 deprecated-1.2.14 dirtyjson-1.0.8 distro-1.9.0 exceptiongroup-1.2.1 feedfinder2-0.0.4 feedparser-6.0.11 filelock-3.14.0 frozenlist-1.4.1 fsspec-2024.6.0 greenlet-3.0.3 h11-0.14.0 html2text-2020.1.16 httpcore-1.0.5 httpx-0.27.0 jieba3k-0.35.1 joblib-1.4.2 llama-index-0.10.43 llama-index-agent-openai-0.2.7 llama-index-cli-0.1.12 llama-index-core-0.10.43 llama-index-embeddings-openai-0.1.10 llama-index-indices-managed-llama-cloud-0.1.6 llama-index-legacy-0.9.48 llama-index-llms-openai-0.1.22 llama-index-multi-modal-llms-openai-0.1.6 llama-index-program-openai-0.1.6 llama-index-question-gen-openai-0.1.3 llama-index-readers-file-0.1.23 llama-index-readers-llama-parse-0.1.4 llama-index-readers-web-0.1.18 llama-index-vector-stores-tidbvector-0.1.2 llama-parse-0.4.4 llamaindex-py-client-0.1.19 lxml-5.2.2 marshmallow-3.21.3 multidict-6.0.5 mypy-extensions-1.0.0 nest-asyncio-1.6.0 networkx-3.2.1 newspaper3k-0.2.8 nltk-3.8.1 numpy-1.26.4 openai-1.33.0 outcome-1.3.0.post0 packaging-24.0 pandas-2.2.2 peewee-3.17.5 pillow-10.3.0 playwright-1.44.0 pydantic-2.7.3 pydantic-core-2.18.4 pyee-11.1.0 pymysql-1.1.1 pypdf-4.2.0 python-dateutil-2.9.0.post0 regex-2024.5.15 requests-2.32.3 requests-file-2.1.0 selenium-4.21.0 sgmllib3k-1.0.0 sniffio-1.3.1 sortedcontainers-2.4.0 soupsieve-2.5 spider-client-0.0.11 sqlalchemy-2.0.30 striprtf-0.0.26 tenacity-8.3.0 tidb-vector-0.0.9 tiktoken-0.7.0 tinysegmenter-0.3 tldextract-5.1.2 tqdm-4.66.4 trio-0.25.1 trio-websocket-0.11.1 typing-extensions-4.12.2 typing-inspect-0.9.0 tzdata-2024.1 wrapt-1.16.0 wsproto-1.2.0 yarl-1.9.4 ``` + TiDB をベクトル ストアとして利用し、効率的なベクトル インデックスによってベクトルを取得するために、SQL を直接使用する必要はありません。TiDB は、Python と次のフレームワークをサポートしています: TiDB Vector Client、SQLAlchemy、Peewee、Django、LangChain、および LlamaIndex このデモでは、[TiDB を LlamaIndex のベクトル ストアとして使用](https://docs.llamaindex.ai/en/stable/examples/vector_stores/TiDBVector/)。 ### 手順 7. シンプルな RAG チャットボット アプリケーションを作成する + `tidb-chat` プログラムを作成します。TiDB に接続し、テーブル `semantic_search_test.llama_index_rag_test` をベクトル ストアとして使用することに注意してください。また、ステップ 2 が正常に完了している必要があります。 ```bash cat > tidb-chat.py <<'EOF' import os import click from sqlalchemy import URL from llama_index.core import VectorStoreIndex, StorageContext from llama_index.vector_stores.tidbvector import TiDBVectorStore from llama_index.readers.web import SimpleWebPageReader tidb_connection_url = URL( "mysql+pymysql", username='serverless_username', password='serverless_password', host='serverless_host', port=serverless_port, database="semantic_search_test", query={"ssl_verify_cert": True, "ssl_verify_identity": True}, ) tidbvec = TiDBVectorStore( connection_string=tidb_connection_url, table_name="llama_index_rag_test", distance_strategy="cosine", vector_dimension=1536, drop_existing_table=False, ) tidb_vec_index = VectorStoreIndex.from_vector_store(tidbvec) storage_context = StorageContext.from_defaults(vector_store=tidbvec) query_engine = tidb_vec_index.as_query_engine(streaming=True) @click.command() def start_chat(): while True: question = click.prompt("Enter your question") response = query_engine.query(question) click.echo(response) if __name__ == '__main__': start_chat() EOF ls -l tidb-chat.py ``` ```output -rw-r--r--. 1 rocky rocky 1157 Jun 20 14:57 tidb-chat.py ``` ### 手順 8. シンプルな RAG チャットボット アプリケーションを実行する + `tidb-chat` プログラムを実行して楽しみましょう。 ```bash python tidb-chat.py ``` ```output Enter your question: ``` + 先ほど作成した `tidb-chat` プログラムで次の質問を試してください。 ```bash what is a dm-task.yaml file? ``` ```output Enter your question: what is a dm-task.yaml file? A `dm-task.yaml` file is a configuration file used for defining the settings and parameters required for a migration task in a DM (Data Migration) cluster. This file contains the necessary information such as the data source, target database, replication rules, filtering criteria, and other configurations essential for executing a migration task. It serves as a blueprint for DM to understand how to handle the data migration process effectively. ``` ```bash How to migrate to TiDB from MySQL? ``` ```output Enter your question: How to migrate to TiDB from MySQL? To migrate to TiDB from MySQL, you can follow these steps: 1. Export all data from MySQL. 2. Import full data to TiDB. 3. Replicate incremental data to TiDB by adding the data source, adding a replication task, checking the migration task status, monitoring the task, and viewing logs. ``` + 時間があれば、さらに質問してみてください。Ctrl + C を押して OS プロンプトに戻ります。 ### 手順 9. 空のデータベースでもう一度やり直す + 1 つのステップですべてを最初からやり直します。このステップでは、TiDB 環境内に空のデータベースを作成してプロセスを開始します。その後、LlamaIndex フレームワークを活用して、指定された URL から直接ベクトルトアを自動的に構築します。この自動化により、ベクトル ストアの作成が効率化されます。 + リモート Linux で、ベクトル ストアの目的を果たすために、TiDB に `chatdb` という名前の新しいデータベースをプロビジョニングします。次に、OS プロンプトに戻ります。 ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt ``` ```sql DROP DATABASE IF EXISTS chatdb; CREATE DATABASE chatdb; EXIT; ``` + `chatdb` をベクトル ストアとして使用するための新しい Python チャット プログラム `tidb-chat-v2.py` を作成します。コードを確認して、`URL()` コンストラクターで使用されるデータベースが `chatdb` であることを確認します。実行中、プログラムはベクトル ストアとしてテーブル `llama_index_rag_test` を作成し、ドキュメントと対応する埋め込みをそこに入力します。 ```bash cat > tidb-chat-v2.py <<'EOF' import os import click from sqlalchemy import URL from llama_index.core import VectorStoreIndex, StorageContext from llama_index.vector_stores.tidbvector import TiDBVectorStore from llama_index.readers.web import SimpleWebPageReader tidb_connection_url = URL( "mysql+pymysql", username='serverless_username', password='serverless_password', host='serverless_host', port=serverless_port, database='chatdb', query={"ssl_verify_cert": True, "ssl_verify_identity": True}, ) tidbvec = TiDBVectorStore( connection_string=tidb_connection_url, table_name="llama_index_rag_test", distance_strategy="cosine", vector_dimension=1536, drop_existing_table=True, ) tidb_vec_index = VectorStoreIndex.from_vector_store(tidbvec) storage_context = StorageContext.from_defaults(vector_store=tidbvec) query_engine = tidb_vec_index.as_query_engine(streaming=True) def prepare_data(url): documents = SimpleWebPageReader(html_to_text=True).load_data([url,]) tidb_vec_index.from_documents(documents, storage_context=storage_context, show_progress=True) urls = [ 'https://docs.pingcap.com/tidb/stable/overview', 'https://docs.pingcap.com/tidb/stable/quick-start-with-tidb', 'https://docs.pingcap.com/tidb/stable/quick-start-with-htap', 'https://docs.pingcap.com/tidb/stable/basic-sql-operations', 'https://docs.pingcap.com/tidb/stable/explore-htap', 'https://docs.pingcap.com/tidb/stable/migration-overview', 'https://docs.pingcap.com/tidb/stable/migration-tools', 'https://docs.pingcap.com/tidb/stable/data-import-best-practices', 'https://docs.pingcap.com/tidb/stable/migrate-aurora-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-small-mysql-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-large-mysql-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-small-mysql-shards-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-large-mysql-shards-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-from-mariadb', 'https://docs.pingcap.com/tidb/stable/migrate-from-csv-files-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-from-sql-files-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-from-parquet-files-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-from-tidb-to-tidb', 'https://docs.pingcap.com/tidb/stable/migrate-from-tidb-to-mysql', ] def start_chat(urls): for url in urls: prepare_data(url) while True: question = click.prompt("Enter your question") response = query_engine.query(question) click.echo(response) if __name__ == '__main__': start_chat(urls) EOF ``` + 新しいバージョンの chat アプリケーションを実行します。 ```bash python tidb-chat-v2.py ``` ```output Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 128.14it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 5.69it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 64.40it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 12.48it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 98.69it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 12.08it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 159.58it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 4.51it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 99.84it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 5.01it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 176.32it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 7.16it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 140.53it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 10.30it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 84.42it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 8.75it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 59.69it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 11.87it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 120.75it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 9.85it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 58.71it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 10.82it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 78.74it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 6/6 [00:00<00:00, 9.22it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 45.15it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 9/9 [00:00<00:00, 15.58it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 72.00it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 14.08it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 94.23it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 11.66it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 130.31it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 9.57it/s] Parsing nodes: 100%|██████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 108.44it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 9.51it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 76.71it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 6/6 [00:00<00:00, 10.05it/s] Parsing nodes: 100%|███████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 95.09it/s] Generating embeddings: 100%|███████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 11.96it/s] ``` ```output Enter your question: ``` + 時間があれば、いくつか質問してみてください。テストが完了したら、Ctrl + C を押して OS プロンプトに戻ります。 ### 手順 10. データベースに入力されたデータを確認する + 指定された TiDB Cloud Serverless インスタンスに接続し、データベース `chatdb` の下にテーブル `llama_index_rag_test` が作成されたことを確認します。 ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt \ -D chatdb ``` ```sql USE chatdb; SHOW TABLES; ``` ```output +----------------------+ | Tables_in_chatdb | +----------------------+ | llama_index_rag_test | +----------------------+ 1 row in set (0.00 sec) ``` + テーブル スキーマを確認します。列 `embedding` は列 `document` のベクトル埋め込みであり、列 `embedding` には `Cosine Distance` インデックスがあることに注意してください。 ```sql SHOW CREATE TABLE llama_index_rag_test\G ``` ```output *************************** 1. row *************************** Table: llama_index_rag_test Create Table: CREATE TABLE `llama_index_rag_test` ( `id` varchar(36) NOT NULL, `embedding` vector(1536) NOT NULL COMMENT 'hnsw(distance=cosine)', `document` text DEFAULT NULL, `meta` json DEFAULT NULL, `create_time` datetime DEFAULT CURRENT_TIMESTAMP, `update_time` datetime DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP, PRIMARY KEY (`id`) /*T![clustered_index] CLUSTERED */ ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin 1 row in set (0.01 sec) ``` + テーブル `llama_index_rag_test` に列 `document` と `embedding` が入力されていることを確認します。 ```sql SELECT COUNT(DISTINCT document), COUNT(DISTINCT embedding) FROM llama_index_rag_test; ``` ```output +--------------------------+---------------------------+ | COUNT(DISTINCT document) | COUNT(DISTINCT embedding) | +--------------------------+---------------------------+ | 99 | 99 | +--------------------------+---------------------------+ 1 row in set (0.02 sec) ``` + OS プロンプトを終了します。 ```sql EXIT ``` ```output Bye ``` ### モジュール 3 のまとめ まだ終わりではありません。次の練習セッションでは、LLM を使用してナレッジ グラフの構築を支援します。その後、このナレッジ グラフを活用してチャット プログラムを強化し、基本的な RAG から Graph RAG に移行します。 ## モジュール 4: シンプルな GraphRAG チャットボットの構築 これまでの練習を通して、RAG を使用して外部の知識ソースから関連情報を取得すると、大規模な言語モデルがこれまで見たことのないドキュメントコレクションからの質問に答えられるようになることを学習しました。ただし、RAG は、テキストコーパス全体に向けられた幅広い質問を処理するには不十分です。ナレッジグラフを使用すると、RAG アプリケーションを単純な RAG から Graph RAG にアップグレードして、LLM プロンプトの強化のために最も類似したテキストを見つけるのではなく、最も関連性の高い知識を取得できます。 この演習では、`gpt-4o` をサポートする OpenAI API キーが必要です。 ![48%](https://lab-static.pingcap.com/quick-demo/TiDBAsGraphStoreForRAGDetails_aa_001_20240610.png) ### ヒント + この演習では、Graph RAG アプローチのオープンソースの Python ベースの実装を使用して、TiDB データベース `chatdb` の Wiki ページから「TiDB」に関するナレッジ グラフを構築します。 + Python ベースの実装では、グラフを表すために TiDB の 2 つのテーブル `entities` と `relationships` を使用します。ナレッジドキュメントとそのベクトル埋め込みは `entities` テーブルに保存されます。 + 埋め込みベースのベクトル検索はグラフへのエントリ ポイントとして機能し、`JOIN` 操作はグラフ内をトラバースして最も関連性の高い知識ノードを取得します。 ### 手順 1. TiDB Cloud Serverless データベースに接続する + 指定された TiDB Cloud Serverless インスタンスに接続し、`chatdb` というデータベースがあることを確認します。 ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt ``` ```sql DROP DATABASE IF EXISTS chatdb; CREATE DATABASE chatdb; EXIT ``` ### 手順 2. アプリケーション環境を準備する + 必要な Python モジュールをインストールします。 ```command pip install \ click==8.1.7 \ PyMySQL==1.1.0 \ SQLAlchemy==2.0.30 \ tidb-vector==0.0.14 \ pydantic==2.7.1 pydantic_core==2.18.2 \ dspy-ai==2.4.9 \ langchain-community==0.2.0 \ wikipedia==1.4.0 \ pyvis==0.3.1 \ openai==1.27.0 \ httpx==0.27.2 ``` ```output ... Successfully installed Mako-1.3.6 PyMySQL-1.1.0 SQLAlchemy-2.0.30 aiohappyeyeballs-2.4.3 aiohttp-3.11.0 aiosignal-1.3.1 alembic-1.14.0 annotated-types-0.7.0 anyio-4.6.2.post1 asttokens-2.4.1 async-timeout-4.0.3 backoff-2.2.1 beautifulsoup4-4.12.3 certifi-2024.8.30 click-8.1.7 colorlog-6.9.0 dataclasses-json-0.6.7 datasets-2.14.7 decorator-5.1.1 dill-0.3.7 distro-1.9.0 dspy-ai-2.4.9 exceptiongroup-1.2.2 executing-2.1.0 filelock-3.16.1 frozenlist-1.5.0 fsspec-2023.10.0 greenlet-3.1.1 h11-0.14.0 httpcore-1.0.6 httpx-0.27.2 huggingface-hub-0.26.2 ipython-8.18.1 jedi-0.19.2 joblib-1.3.2 jsonpatch-1.33 jsonpickle-4.0.0 langchain-0.2.17 langchain-community-0.2.0 langchain-core-0.2.43 langchain-text-splitters-0.2.4 langsmith-0.1.143 marshmallow-3.23.1 matplotlib-inline-0.1.7 multidict-6.1.0 multiprocess-0.70.15 mypy-extensions-1.0.0 networkx-3.2.1 numpy-1.26.4 openai-1.27.0 optuna-4.1.0 orjson-3.10.11 packaging-24.2 pandas-2.2.3 parso-0.8.4 pexpect-4.9.0 prompt-toolkit-3.0.48 propcache-0.2.0 ptyprocess-0.7.0 pure-eval-0.2.3 pyarrow-18.0.0 pyarrow-hotfix-0.6 pydantic-2.7.1 pydantic-core-2.18.2 pygments-2.18.0 python-dateutil-2.9.0.post0 pyvis-0.3.1 regex-2024.11.6 requests-toolbelt-1.0.0 sniffio-1.3.1 soupsieve-2.6 stack-data-0.6.3 structlog-24.4.0 tenacity-8.5.0 tidb-vector-0.0.14 tqdm-4.67.0 traitlets-5.14.3 typing-extensions-4.12.2 typing-inspect-0.9.0 tzdata-2024.2 ujson-5.10.0 wcwidth-0.2.13 wikipedia-1.4.0 xxhash-3.5.0 yarl-1.17.1 ``` + 環境に[OpenAI APIキー](https://help.openai.com/en/articles/4936850-where-do-i-find-my-openai-api-key)を設定します。`${OPENAI_API_KEY}`を実際の値に置き換え、リモート Linux で以下のコマンドを実行します。前の練習で設定を完了している場合は、この手順をスキップできます。 ```bash export OPENAI_API_KEY=${OPENAI_API_KEY} ``` ### 手順 3. チャットボットのナレッジ グラフを構築する + 「TiDB」の Wikipedia Loader クエリからナレッジグラフを構築します。 + `build-graph.py` という名前の Python スクリプトを作成するには、次のコマンドを実行します。下のコード領域をクリックすると、`cat` コマンド全体を簡単にコピーできます。 ```bash cat > build-graph.py <<'EOF' import os import pymysql import dspy import enum import openai import tidb_vector from pymysql import Connection from pymysql.cursors import DictCursor from dspy.functional import TypedPredictor from pydantic import BaseModel, Field from typing import Mapping, Any, Optional, List from langchain_community.document_loaders import WikipediaLoader from pyvis.network import Network from IPython.display import HTML from sqlalchemy import ( Column, Integer, String, Text, JSON, ForeignKey, BLOB, Enum as SQLEnum, DateTime, URL, create_engine, or_, ) from sqlalchemy.orm import relationship, Session, declarative_base, joinedload from tidb_vector.sqlalchemy import VectorType, VectorAdaptor class Entity(BaseModel): """List of entities extracted from the text to form the knowledge graph""" name: str = Field( description="Name of the entity, it should be a clear and concise term" ) description: str = Field( description=( "Description of the entity, it should be a complete and comprehensive sentence, not few words. " "Sample description of entity 'TiDB in-place upgrade': " "'Upgrade TiDB component binary files to achieve upgrade, generally use rolling upgrade method'" ) ) class Relationship(BaseModel): """List of relationships extracted from the text to form the knowledge graph""" source_entity: str = Field( description="Source entity name of the relationship, it should an existing entity in the Entity list" ) target_entity: str = Field( description="Target entity name of the relationship, it should an existing entity in the Entity list" ) relationship_desc: str = Field( description=( "Description of the relationship, it should be a complete and comprehensive sentence, not few words. " "Sample relationship description: 'TiDB will release a new LTS version every 6 months.'" ) ) class KnowledgeGraph(BaseModel): """Graph representation of the knowledge for text.""" entities: List[Entity] = Field( description="List of entities in the knowledge graph" ) relationships: List[Relationship] = Field( description="List of relationships in the knowledge graph" ) class ExtractGraphTriplet(dspy.Signature): text = dspy.InputField( desc="a paragraph of text to extract entities and relationships to form a knowledge graph" ) knowledge: KnowledgeGraph = dspy.OutputField( desc="Graph representation of the knowledge extracted from the text." ) class Extractor(dspy.Module): def __init__(self): super().__init__() self.prog_graph = TypedPredictor(ExtractGraphTriplet) def forward(self, text): return self.prog_graph( text=text, config={ "response_format": {"type": "json_object"}, }, ) def interactive_graph(kg: KnowledgeGraph) -> str: net = Network(notebook=True, cdn_resources="remote") node_map = {} for index in range(len(kg.entities)): node_map[kg.entities[index].name] = index net.add_node( index, label=kg.entities[index].name, title=kg.entities[index].description ) for index in range(len(kg.relationships)): relation = kg.relationships[index] src_index = node_map[relation.source_entity] target_index = node_map[relation.target_entity] net.add_edge(src_index, target_index) filename = "kg.html" net.save_graph(filename) return filename def get_query_embedding(query: str): open_ai_client = openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY")) response = open_ai_client.embeddings.create( input=[query], model="text-embedding-3-small" ) return response.data[0].embedding def generate_result(query: str, entities, relationships): open_ai_client = openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY")) entities_prompt = "\n".join( map(lambda e: f'(Name: "{e.name}", Description: "{e.description}")', entities) ) relationships_prompt = "\n".join( map(lambda r: f'"{r.relationship_desc}"', relationships) ) response = open_ai_client.chat.completions.create( model="gpt-4o", messages=[ { "role": "system", "content": "Please carefully think the user's " + "question and ONLY use the content below to generate answer:\n" + f"Entities: {entities_prompt}, Relationships: {relationships_prompt}", }, {"role": "user", "content": query}, ], ) return response.choices[0].message.content def get_db_url(): return URL( drivername="mysql+pymysql", username="serverless_username", password="serverless_password", host="serverless_host", port=serverless_port, database="chatdb", query={"ssl_verify_cert": True, "ssl_verify_identity": True}, ) engine = create_engine(get_db_url(), pool_recycle=300) Base = declarative_base() class DatabaseEntity(Base): id = Column(Integer, primary_key=True) name = Column(String(512)) description = Column(Text) description_vec = Column(VectorType(1536)) __tablename__ = "entities" class DatabaseRelationship(Base): id = Column(Integer, primary_key=True) source_entity_id = Column(Integer, ForeignKey("entities.id")) target_entity_id = Column(Integer, ForeignKey("entities.id")) relationship_desc = Column(Text) source_entity = relationship("DatabaseEntity", foreign_keys=[source_entity_id]) target_entity = relationship("DatabaseEntity", foreign_keys=[target_entity_id]) __tablename__ = "relationships" def clean_knowledge_graph(kg: KnowledgeGraph) -> KnowledgeGraph: entity_name_set = set(map(lambda e: e.name, kg.entities)) kg.relationships = list( filter( lambda r: r.source_entity in entity_name_set and r.target_entity in entity_name_set, kg.relationships, ) ) return kg def save_knowledge_graph(kg: KnowledgeGraph): data_entities = list( map( lambda e: DatabaseEntity( name=e.name, description=e.description, description_vec=get_query_embedding(e.description), ), kg.entities, ) ) with Session(engine) as session: session.add_all(data_entities) session.flush() entity_id_map = dict(map(lambda e: (e.name, e.id), data_entities)) print(entity_id_map) data_relationships = list( map( lambda r: DatabaseRelationship( source_entity_id=entity_id_map[r.source_entity], target_entity_id=entity_id_map[r.target_entity], relationship_desc=r.relationship_desc, ), kg.relationships, ) ) session.add_all(data_relationships) session.commit() def retrieve_entities_relationships(question_embedding) -> (List[DatabaseEntity], List[DatabaseRelationship]): with Session(engine) as session: entity = ( session.query(DatabaseEntity) .order_by( DatabaseEntity.description_vec.cosine_distance(question_embedding) ) .limit(1) .first() ) entities = {entity.id: entity} relationships = ( session.query(DatabaseRelationship) .options( joinedload(DatabaseRelationship.source_entity), joinedload(DatabaseRelationship.target_entity), ) .filter( or_( DatabaseRelationship.source_entity == entity, DatabaseRelationship.target_entity == entity, ) ) ) for r in relationships: entities.update( { r.source_entity.id: r.source_entity, r.target_entity.id: r.target_entity, } ) return entities.values(), relationships extractor = Extractor() Base.metadata.create_all(engine) # Create Index VectorAdaptor(engine).create_vector_index( DatabaseEntity.description_vec, tidb_vector.DistanceMetric.COSINE, skip_existing=True # For L2 distance, use tidb_vector.DistanceMetric.L2 ) open_ai_client = dspy.OpenAI( model="gpt-4o", api_key=os.getenv("OPENAI_API_KEY"), max_tokens=4096 ) dspy.settings.configure(lm=open_ai_client) wiki = WikipediaLoader(query="TiDB").load() pred = extractor(text=wiki[0].page_content) knowledge_graph = clean_knowledge_graph(pred.knowledge) interactive_graph(knowledge_graph) save_knowledge_graph(knowledge_graph) EOF ls -l build-graph.py ``` ```output -rw-r--r--. 1 rocky rocky 8777 Jun 20 15:02 build-graph.py ``` + Python スクリプト `build-graph.py` を実行してグラフを構築します。データベース `chatdb` の下に `entities` および `relationships` テーブルが作成されます。スクリプトが完了するまで待ちます。約 2 分かかります。出力は、表示されている出力と必ずしも同じではありません。 ```bash python build-graph.py ``` ```output {'TiDB': 90001, 'PingCAP': 90002, 'Apache 2.0': 90003, "Google's Spanner and F1 papers": 90004, 'TiDB 7.5 GA': 90005, 'TiDB 7.1 GA': 90006, 'TiDB 6.0 GA': 90007, 'TiDB 5.0 GA': 90008, 'TiDB 4.0 GA': 90009, 'TiDB 3.0 GA': 90010, 'TiDB 2.0 GA': 90011, 'TiDB 1.0 GA': 90012, 'Horizontal scalability': 90013, 'MySQL compatibility': 90014, 'Distributed transactions with strong consistency': 90015, 'Cloud native': 90016, 'TiKV': 90017, 'Real-time HTAP': 90018, 'TiFlash': 90019, 'High availability': 90020, 'Raft consensus algorithm': 90021, 'Kubernetes with Operator': 90022, 'TiDB Operator': 90023, 'TiUP': 90024, 'TiDB Ansible': 90025, 'Docker': 90026, 'TiDB Data Migration (DM)': 90027, 'Backup & Restore (BR)': 90028, 'Dumpling': 90029} ``` ### 手順 4. 構築したナレッジ グラフを確認する + TiDB に接続し、構築したシンプルなナレッジ グラフを確認します。 + `chatdb` データベースに接続します。 ```bash mysql \ --comments -u 'serverless_username' -p'serverless_password' \ -h serverless_host -P serverless_port \ --ssl-mode=VERIFY_IDENTITY --ssl-ca=/etc/pki/tls/certs/ca-bundle.crt \ -D chatdb ``` ```output mysql: [Warning] Using a password on the command line interface can be insecure. Welcome to the MySQL monitor. Commands end with ; or \g. Your MySQL connection id is 2168969 Server version: 5.7.28-TiDB-Serverless TiDB Server (Apache License 2.0) Community Edition, MySQL 5.7 compatible Copyright (c) 2000, 2024, Oracle and/or its affiliates. Oracle is a registered trademark of Oracle Corporation and/or its affiliates. Other names may be trademarks of their respective owners. Type 'help;' or '\h' for help. Type '\c' to clear the current input statement. mysql> ``` + Python スクリプトによって生成されたテーブル `chatdb.entities` を確認します。HNSW インデックスが列 `description_vec` で宣言されていることに注意してください。出力は、表示された出力と必ずしも同一ではないことに注意してください。 ```sql SHOW CREATE TABLE entities\G ``` ```output *************************** 1. row *************************** Table: entities Create Table: CREATE TABLE `entities` ( `id` int(11) NOT NULL AUTO_INCREMENT, `name` varchar(512) DEFAULT NULL, `description` text DEFAULT NULL, `description_vec` vector(1536) DEFAULT NULL COMMENT 'HNSW(distance=cosine)', PRIMARY KEY (`id`) /*T![clustered_index] CLUSTERED */, VECTOR INDEX `vec_idx_description_vec` ((VEC_COSINE_DISTANCE(`description_vec`))) ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin AUTO_INCREMENT=60001 1 row in set (0.02 sec) ``` ```sql SELECT id, name, SUBSTR(description, 1,30) AS desc_head, VEC_DIMS(description_vec) FROM entities ORDER BY name; ``` ```output +----+--------------------------------------------------+--------------------------------+---------------------------+ | id | name | desc_head | VEC_DIMS(description_vec) | +----+--------------------------------------------------+--------------------------------+---------------------------+ | 3 | Apache 2.0 | Apache 2.0 is the license unde | 1536 | | 26 | Backup & Restore (BR) | Backup & Restore (BR) is a dis | 1536 | | 31 | Cloud Native Computing Foundation (CNCF) | The Cloud Native Computing Fou | 1536 | | 18 | Cloud native | TiDB is designed to work in th | 1536 | | 17 | Distributed transactions with strong consistency | TiDB internally shards a table | 1536 | | 24 | Docker | Docker can be used to deploy T | 1536 | | 4 | Google's Spanner and F1 papers | Google's Spanner and F1 papers | 1536 | | 20 | High availability | TiDB uses the Raft consensus a | 1536 | | 15 | Horizontal scalability | TiDB can expand both SQL proce | 1536 | | 21 | Kubernetes with Operator | TiDB can be deployed in a Kube | 1536 | | 16 | MySQL compatibility | TiDB acts like it is a MySQL 8 | 1536 | | 2 | PingCAP | PingCAP is the primary develop | 1536 | | 29 | Raft consensus algorithm | The Raft consensus algorithm i | 1536 | | 19 | Real-time HTAP | TiDB supports both online tran | 1536 | | 1 | TiDB | TiDB is an open-source NewSQL | 1536 | | 14 | TiDB 1.0 GA | TiDB 1.0 GA was released on Oc | 1536 | | 13 | TiDB 2.0 GA | TiDB 2.0 GA was released on Ap | 1536 | | 12 | TiDB 3.0 GA | TiDB 3.0 GA was released on Ju | 1536 | | 11 | TiDB 4.0 GA | TiDB 4.0 GA was released on Ma | 1536 | | 10 | TiDB 5.0 GA | TiDB 5.0 GA was released on Ap | 1536 | | 9 | TiDB 6.0 GA | TiDB 6.0 GA was released on Ap | 1536 | | 8 | TiDB 7.1 GA | TiDB 7.1 GA was released on Ma | 1536 | | 7 | TiDB 7.5 GA | TiDB 7.5 GA was released on De | 1536 | | 6 | TiDB 8.1 GA | TiDB 8.1 GA was released on Ma | 1536 | | 5 | TiDB 8.5 GA | TiDB 8.5 GA was released on De | 1536 | | 23 | TiDB Ansible | TiDB can be deployed using Ans | 1536 | | 25 | TiDB Data Migration (DM) | TiDB Data Migration (DM) is su | 1536 | | 30 | TiDB Operator | TiDB Operator is a method of p | 1536 | | 28 | TiFlash | TiFlash is a columnstore stora | 1536 | | 27 | TiKV | TiKV is the storage layer of T | 1536 | | 22 | TiUP | TiDB 4.0 introduces TiUP, a cl | 1536 | +----+--------------------------------------------------+--------------------------------+---------------------------+ 31 rows in set (0.34 sec) ``` + テーブル `chatdb.relationships` スキーマを確認し、テーブル `entities` と `relationshipts` の結合から関係を取得します。出力は、表示された出力と必ずしも同一ではないことに注意してください。 ```sql SHOW CREATE TABLE relationships\G ``` ```output *************************** 1. row *************************** Table: relationships Create Table: CREATE TABLE `relationships` ( `id` int(11) NOT NULL AUTO_INCREMENT, `source_entity_id` int(11) DEFAULT NULL, `target_entity_id` int(11) DEFAULT NULL, `relationship_desc` text DEFAULT NULL, PRIMARY KEY (`id`) /*T![clustered_index] CLUSTERED */, KEY `fk_1` (`source_entity_id`), KEY `fk_2` (`target_entity_id`), CONSTRAINT `fk_1` FOREIGN KEY (`source_entity_id`) REFERENCES `chatdb`.`entities` (`id`), CONSTRAINT `fk_2` FOREIGN KEY (`target_entity_id`) REFERENCES `chatdb`.`entities` (`id`) ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin AUTO_INCREMENT=60001 1 row in set (0.00 sec) ``` ```sql SELECT source.name AS src_name, SUBSTR(r.relationship_desc, 1, 50) relation_head, target.name AS tgt_name FROM entities AS source, entities AS target, relationships as r WHERE source.id = r.source_entity_id AND target.id = r.target_entity_id ORDER BY src_name; ``` ```output +--------------------------+----------------------------------------------------+--------------------------------------------------+ | src_name | relation_head | tgt_name | +--------------------------+----------------------------------------------------+--------------------------------------------------+ | Raft consensus algorithm | The Raft consensus algorithm is used by TiDB to en | High availability | | TiDB | TiDB acts like it is a MySQL 8.0 server to applica | MySQL compatibility | | TiDB | TiDB supports both online transaction processing ( | Real-time HTAP | | TiDB | Dumpling is a data export tool for TiDB. | Dumpling | | TiDB | Backup & Restore (BR) is a distributed backup and | Backup & Restore (BR) | | TiDB | TiDB Data Migration (DM) is a tool for replicating | TiDB Data Migration (DM) | | TiDB | Docker can be used to deploy TiDB in a containeriz | Docker | | TiDB | TiDB can be deployed using Ansible with a TiDB Ans | TiDB Ansible | | TiDB | TiUP is a cluster operation and maintenance tool i | TiUP | | TiDB | TiDB can be deployed in a Kubernetes-enabled cloud | Kubernetes with Operator | | TiDB | TiDB uses the Raft consensus algorithm to ensure d | High availability | | TiDB | TiDB can expand both SQL processing and storage ca | Horizontal scalability | | TiDB | TiDB acts like it is a MySQL 8.0 server to applica | MySQL compatibility | ... | TiDB 5.0 GA | TiDB 5.0 GA was released on April 7, 2021. | TiDB | | TiDB 5.0 GA | TiDB 5.0 GA was released on April 7, 2021. | TiDB | | TiDB 6.0 GA | TiDB 6.0 GA was released on April 7, 2022. | TiDB | | TiDB 6.0 GA | TiDB 6.0 GA was released on April 7, 2022. | TiDB | | TiDB 6.0 GA | TiDB 6.0 GA was released on April 7, 2022. | TiDB | | TiDB 7.1 GA | TiDB 7.1 GA was released on May 31, 2023. | TiDB | | TiDB 7.1 GA | TiDB 7.1 GA was released on May 31, 2023. | TiDB | | TiDB 7.1 GA | TiDB 7.1 GA was released on May 31, 2023. | TiDB | | TiDB 7.5 GA | TiDB 7.5 GA was released on December 1, 2023. | TiDB | | TiDB 7.5 GA | TiDB 7.5 GA was released on December 1, 2023. | TiDB | | TiDB 7.5 GA | TiDB 7.5 GA was released on December 1, 2023. | TiDB | | TiDB Operator | TiDB Operator was originally developed by PingCAP | PingCAP | | TiDB Operator | TiDB Operator was originally developed by PingCAP | PingCAP | | TiDB Operator | TiDB Operator is a method of packaging, deploying, | Kubernetes with Operator | | TiFlash | TiFlash is a columnstore storage engine used by Ti | Real-time HTAP | | TiKV | TiKV, the storage layer of TiDB, became a Cloud Na | Cloud native | +--------------------------+----------------------------------------------------+--------------------------------------------------+ 81 rows in set (0.01 sec) ``` + リモート Linux 上の OS プロンプトを終了します。 ```sql EXIT ``` ### 手順 5. GraphRAG チャットボット プログラムを作成する + 基本的な Graph RAG アプリケーションを使用して、ナレッジ グラフの動作を実演します。結果を練習 3 と比較します。 + Python スクリプト `test-graph.py` を作成します。 ```bash cat > test-graph.py <<'EOF' import openai import os import click from sqlalchemy import ( Column, Integer, String, Text, ForeignKey, URL, create_engine, or_, ) from typing import Mapping, Any, Optional, List from sqlalchemy.orm import relationship, Session, declarative_base, joinedload from tidb_vector.sqlalchemy import VectorType def get_db_url(): return URL( drivername="mysql+pymysql", username="serverless_username", password="serverless_password", host="serverless_host", port=serverless_port, database="chatdb", query={"ssl_verify_cert": True, "ssl_verify_identity": True}, ) engine = create_engine(get_db_url(), pool_recycle=300) Base = declarative_base() class DatabaseEntity(Base): id = Column(Integer, primary_key=True) name = Column(String(512)) description = Column(Text) description_vec = Column(VectorType(1536)) __tablename__ = "entities" class DatabaseRelationship(Base): id = Column(Integer, primary_key=True) source_entity_id = Column(Integer, ForeignKey("entities.id")) target_entity_id = Column(Integer, ForeignKey("entities.id")) relationship_desc = Column(Text) source_entity = relationship("DatabaseEntity", foreign_keys=[source_entity_id]) target_entity = relationship("DatabaseEntity", foreign_keys=[target_entity_id]) __tablename__ = "relationships" def get_query_embedding(query: str): open_ai_client = openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY")) response = open_ai_client.embeddings.create( input=[query], model="text-embedding-3-small" ) return response.data[0].embedding def retrieve_entities_relationships( question_embedding, ) -> (List[DatabaseEntity], List[DatabaseRelationship]): with Session(engine) as session: entity = ( session.query(DatabaseEntity) .order_by( DatabaseEntity.description_vec.cosine_distance(question_embedding) ) .limit(1) .first() ) entities = {entity.id: entity} relationships = ( session.query(DatabaseRelationship) .options( joinedload(DatabaseRelationship.source_entity), joinedload(DatabaseRelationship.target_entity), ) .filter( or_( DatabaseRelationship.source_entity == entity, DatabaseRelationship.target_entity == entity, ) ) ) for r in relationships: entities.update( { r.source_entity.id: r.source_entity, r.target_entity.id: r.target_entity, } ) return entities.values(), relationships def generate_result(query: str, entities, relationships): open_ai_client = openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY")) entities_prompt = "\n".join( map(lambda e: f'(Name: "{e.name}", Description: "{e.description}")', entities) ) relationships_prompt = "\n".join( map(lambda r: f'"{r.relationship_desc}"', relationships) ) response = open_ai_client.chat.completions.create( model="gpt-4o", messages=[ { "role": "system", "content": "Please carefully think the user's " + "question and ONLY use the content below to generate answer:\n" + f"Entities: {entities_prompt}, Relationships: {relationships_prompt}", }, {"role": "user", "content": query}, ], ) return response.choices[0].message.content @click.command() def start_chat(): while True: question = click.prompt("Enter your question") question_embedding = get_query_embedding(question) entities, relationships = retrieve_entities_relationships(question_embedding) result = generate_result(question, entities, relationships) click.echo(result) if __name__ == '__main__': start_chat() EOF ls -l test-graph.py ``` ```output -rw-r--r--. 1 rocky rocky 4154 Jun 20 15:14 test-graph.py ``` ### 手順 6. GraphRAG チャットボット プログラムを実行する + スクリプト `test-graph.py` を実行します。 ```bash python test-graph.py ``` ```output Enter your question: ``` + 質問 1 を試してください。この知識は TiDB ナレッジ グラフには存在しません。「What is the relationship between raccoon and umbrella?」 ```bash What is the relationship between raccoon and umbrella? ``` ```output Enter your question: What is the relationship between raccoon and umbrella? The provided information does not contain any details about raccoons or umbrellas. Therefore, I can't provide a relationship between these two based on the given content. ``` + 質問 2 に挑戦してください。この知識は TiDB ナレッジ グラフでカバーされています: 「What is the relationship between PD and Raft?」 ```bash What is the relationship between PD and Raft? ``` ```output Enter your question: What is the relationship between PD and Raft? The Raft consensus algorithm is used by TiDB to ensure data availability and replication. However, the relationship specifically between PD and Raft is not detailed in the provided content. Generally speaking, in TiDB, PD (Placement Driver) handles metadata management and scheduling, and works in conjunction with the Raft consensus algorithm to maintain high availability and data consistency through leader election and automatic failover mechanisms. ``` + 質問 3 に答えてください。この知識は、TiDB ナレッジ グラフでカバーされています。「Which tools or solutions are most relevant for migrating to TiDB?」 ``` Which tools or solutions are most relevant for migrating to TiDB? ``` ```output Enter your question: Which tools or solutions are most relevant for migrating to TiDB? The most relevant tool for migrating to TiDB is **TiDB Data Migration (DM)**. This tool is specifically designed for replicating data from sharded MySQL or MariaDB tables to TiDB, facilitating near real-time analytical workloads. ``` + モジュール 3 で尋ねたのと同じ質問を試してください。今回は、単純な RAG ではなく Graph RAG を使用して、2 つの回答を比較します。どちらが適切でしょうか。出力は、表示された出力と必ずしも同一ではないことに注意してください。 ```bash How to migrate to TiDB from MySQL? ``` ```output Enter your question: How to migrate to TiDB from MySQL? To migrate to TiDB from MySQL, you can use TiDB Data Migration (DM). TiDB DM is a tool designed specifically for replicating data from sharded MySQL or MariaDB tables to TiDB, allowing near real-time analytical workloads. Here are the general steps to perform the migration: 1. **Install TiDB DM**: Set up TiDB Data Migration by following the official installation guide provided by PingCAP. 2. **Configure Source Database**: Prepare your MySQL or MariaDB databases with the necessary configurations and privileges required for data replication. 3. **Deploy and Configure DM Workers and Master**: Deploy TiDB DM components, including DM workers and the DM master. Configure these components by defining the source database configurations, replication tasks, and target TiDB cluster details. 4. **Create Data Migration Tasks**: Define and create data migration tasks that detail what data from the source databases should be replicated to TiDB. This includes sharding rules if applicable. 5. **Start Data Migration**: Initiate the data migration tasks to begin replicating data from the MySQL or MariaDB source tables to TiDB. 6. **Monitor and Manage**: Monitor the migration process using provided tools and logs to ensure it progresses as expected. Make adjustments and handle any issues that arise. By following these steps with TiDB Data Migration (DM), you can efficiently migrate your data from MySQL to TiDB while maintaining data integrity and allowing for real-time workload processing on TiDB. For detailed instructions and best practices, refer to the official TiDB documentation. ``` + 時間があれば、さらに質問してみてください。Ctrl + C を押して OS プロンプトに戻ります。 + リモート Linux からログアウトします。 ```bash exit ``` ## 結論 K 近傍 (KNN) 検索は、ベクトル空間内の特定のポイントに対して K 個の最も近いポイントを見つける問題です。この問題を解決する最も簡単な方法は、ベクトル空間内のすべてのポイントと参照ポイント間の距離を計算する総当たり検索です。この方法は完全な精度を保証しますが、通常は実際のアプリケーションには遅すぎます。したがって、近傍検索の問題は、近似アルゴリズムで解決されることがよくあります。 ```hbsptform { "ui": { "titleLevel": 3 }, "portalId": "4466002", "formId": "62bfb32a-a60a-4f9d-b2dc-1107a93db568", "region": "na1", "sfdcCampaignId": "701RC00000NdFlAYAV" } ```