-- =============================================================================
--  ai_rag_chunk
--
--  One row per text chunk in a RAG collection. The embedding column stores
--  the vector as a base64-packed Float32Array (4 bytes/dim). Cosine search
--  is done in PHP after fetching all chunks for the collection — fine up
--  to ~10k chunks per collection. For larger sets, drop in a vector index
--  extension (pgvector, MySQL HeatWave LakeHouse vector) and update
--  Rag::search() accordingly.
-- =============================================================================

CREATE TABLE IF NOT EXISTS ai_rag_chunk (
    ai_rag_chunk_id            INT(11)      NOT NULL AUTO_INCREMENT,
    ai_rag_chunk_collection_id INT(11)      NOT NULL,
    ai_rag_chunk_text          MEDIUMTEXT   NOT NULL,
    ai_rag_chunk_metadata      JSON         NULL,
    ai_rag_chunk_embedding     LONGTEXT     NOT NULL,    -- base64-packed Float32Array
    ai_rag_chunk_token_count   INT(11)      NOT NULL DEFAULT 0,
    ai_rag_chunk_source        VARCHAR(512) NULL,        -- denormalised from metadata for indexing
    ai_rag_chunk_created_at    TIMESTAMP    NOT NULL DEFAULT CURRENT_TIMESTAMP,
    PRIMARY KEY (ai_rag_chunk_id),
    KEY ai_rag_chunk_collection_idx (ai_rag_chunk_collection_id),
    KEY ai_rag_chunk_source_idx (ai_rag_chunk_source),
    CONSTRAINT ai_rag_chunk_collection_fk
        FOREIGN KEY (ai_rag_chunk_collection_id)
        REFERENCES ai_rag_collection (ai_rag_collection_id)
        ON DELETE CASCADE
) ENGINE = InnoDB DEFAULT CHARSET = utf8mb4 COLLATE = utf8mb4_unicode_ci;