RAG精度向上実践: Oracle Autonomous DBを用いたメタデータ活用チャンキング設計

CREATE TABLE rag_chunks (
  chunk_id        NUMBER GENERATED BY DEFAULT AS IDENTITY PRIMARY KEY,
  file_name       VARCHAR2(512),         -- 出典（ファイル/URL）
  content_type    VARCHAR2(50),          -- Text | Table | Diagram
  prev_chunk_id   NUMBER,
  next_chunk_id   NUMBER,
  chunk_text      CLOB,
  metadata        JSON
);

CREATE TABLE rag_embeddings (
  id         NUMBER GENERATED BY DEFAULT AS IDENTITY PRIMARY KEY,
  chunk_id   NUMBER,
  embedding  VECTOR,
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

INSERT INTO RAG_EMBEDDINGS (CHUNK_ID, EMBEDDING, CREATED_AT)
SELECT
   ET.EMBED_ID   AS CHUNK_ID,
   TO_VECTOR(ET.EMBED_VECTOR) AS EMBEDDING,
   CURRENT_TIMESTAMP AS CREATED_AT
FROM
   DOCUMENTATION_TAB DT,
   DBMS_VECTOR_CHAIN.UTL_TO_EMBEDDINGS(
      DBMS_VECTOR_CHAIN.UTL_TO_CHUNKS(
         DBMS_VECTOR_CHAIN.UTL_TO_TEXT(DT.CHUNK_TEXT),
         JSON('{"max": "400", "overlap": "20", "language": "JAPANESE", "normalize": "all"}')
       ),
      JSON('{"provider": "ocigenai", "credential_name": "VECTOR_CRED", "url": "VECTOR_URI", "model": "VECTOR_MODEL"}')
   ) T,
   JSON_TABLE(
      T.COLUMN_VALUE, 
      '$[*]' COLUMNS (
         EMBED_ID    NUMBER PATH '$.embed_id', 
         EMBED_VECTOR CLOB PATH '$.embed_vector'
     )
  ) ET
WHERE DT.ID = :doc_id;

SELECT SOURCE_DOC_ID
FROM {vec_table_name}
ORDER BY VECTOR_DISTANCE(EMBEDDING, TO_VECTOR(:query_embed_vector), COSINE)
FETCH EXACT FIRST :top_k ROWS ONLY

WITH cte AS (
    SELECT
         id,
         file_name,
         metadata,
         text_content,
         LAG(id) OVER (PARTITION BY file_name ORDER BY id)  AS prev_id,
         LEAD(id) OVER (PARTITION BY file_name ORDER BY id) AS next_id
    FROM {doc_table_name}
)
SELECT
     c.id                AS current_id,
     c.metadata          AS current_metadata,
     c.text_content      AS current_text_content,

     p.id                AS prev_id,
     p.metadata          AS prev_metadata,
     p.text_content      AS prev_text_content,

     n.id                AS next_id,
     n.metadata          AS next_metadata,
     n.text_content      AS next_text_content
FROM cte c
LEFT JOIN {doc_table_name} p
ON p.id = c.prev_id
AND p.file_name = c.file_name
LEFT JOIN {doc_table_name} n
ON n.id = c.next_id
AND n.file_name = c.file_name
WHERE c.id = {doc_id}