96 lines
3.9 KiB
MySQL
96 lines
3.9 KiB
MySQL
-- Hybrid retrieval remains database-driven: dense vector similarity handles
|
|
-- paraphrases, while lexical similarity distinguishes decisive request terms.
|
|
MERGE INTO sg_game_scope_policy t
|
|
USING (
|
|
SELECT 'QA_VECTOR_LEXICAL_WEIGHT' AS policy_key, 0.350000 AS number_value,
|
|
'Weight of normalized lexical question similarity in runtime Few-shot reranking.' AS description
|
|
FROM dual
|
|
UNION ALL
|
|
SELECT 'QA_VECTOR_HYBRID_SCORE_MARGIN', 0.050000,
|
|
'Maximum hybrid-score difference from the best runtime Few-shot candidate.'
|
|
FROM dual
|
|
) s
|
|
ON (t.policy_key = s.policy_key)
|
|
WHEN MATCHED THEN UPDATE SET
|
|
t.number_value = s.number_value,
|
|
t.description = s.description,
|
|
t.active_yn = 'Y',
|
|
t.updated_at = SYSTIMESTAMP
|
|
WHEN NOT MATCHED THEN INSERT (
|
|
policy_key, number_value, text_value, description, active_yn
|
|
) VALUES (
|
|
s.policy_key, s.number_value, NULL, s.description, 'Y'
|
|
)
|
|
/
|
|
|
|
CREATE OR REPLACE FUNCTION sg_qa_vector_search(
|
|
p_question IN CLOB,
|
|
p_top_k IN PLS_INTEGER DEFAULT 3,
|
|
p_target_type IN VARCHAR2 DEFAULT 'ANY'
|
|
) RETURN SYS_REFCURSOR AUTHID DEFINER
|
|
IS
|
|
v_query_vector VECTOR;
|
|
v_results SYS_REFCURSOR;
|
|
v_target_type VARCHAR2(16) := UPPER(TRIM(NVL(p_target_type, 'ANY')));
|
|
v_max_cosine_distance NUMBER;
|
|
v_lexical_weight NUMBER;
|
|
v_hybrid_margin NUMBER;
|
|
BEGIN
|
|
IF p_question IS NULL THEN RAISE_APPLICATION_ERROR(-20003, 'question is required.'); END IF;
|
|
IF p_top_k IS NULL OR p_top_k < 1 OR p_top_k > 20 THEN
|
|
RAISE_APPLICATION_ERROR(-20004, 'top_k must be between 1 and 20.');
|
|
END IF;
|
|
IF v_target_type NOT IN ('NONE', 'SINGLE', 'MULTI', 'ALL', 'ANY') THEN
|
|
RAISE_APPLICATION_ERROR(-20005, 'invalid target type.');
|
|
END IF;
|
|
|
|
SELECT number_value INTO v_max_cosine_distance FROM sg_game_scope_policy
|
|
WHERE policy_key = 'QA_VECTOR_MAX_COSINE_DISTANCE' AND active_yn = 'Y';
|
|
SELECT number_value INTO v_lexical_weight FROM sg_game_scope_policy
|
|
WHERE policy_key = 'QA_VECTOR_LEXICAL_WEIGHT' AND active_yn = 'Y';
|
|
SELECT number_value INTO v_hybrid_margin FROM sg_game_scope_policy
|
|
WHERE policy_key = 'QA_VECTOR_HYBRID_SCORE_MARGIN' AND active_yn = 'Y';
|
|
|
|
v_query_vector := DBMS_VECTOR.UTL_TO_EMBEDDING(
|
|
p_question, JSON(sg_qa_vector_params('search_query'))
|
|
);
|
|
|
|
OPEN v_results FOR
|
|
SELECT example_id, question, answer_sql, answer_text, embedding_model,
|
|
reference_kind, target_type, object_role, source_case_id, source_type,
|
|
cosine_distance
|
|
FROM (
|
|
SELECT s.*,
|
|
MAX(s.hybrid_score) OVER () AS best_hybrid_score
|
|
FROM (
|
|
SELECT c.*,
|
|
((1 - v_lexical_weight) * (1 - c.cosine_distance)
|
|
+ v_lexical_weight * c.lexical_similarity) AS hybrid_score
|
|
FROM (
|
|
SELECT example_id, question, answer_sql, answer_text, embedding_model,
|
|
reference_kind, target_type, object_role, source_case_id, source_type,
|
|
VECTOR_DISTANCE(embedding, v_query_vector, COSINE) AS cosine_distance,
|
|
UTL_MATCH.JARO_WINKLER_SIMILARITY(
|
|
DBMS_LOB.SUBSTR(question, 4000, 1),
|
|
DBMS_LOB.SUBSTR(p_question, 4000, 1)
|
|
) / 100 AS lexical_similarity
|
|
FROM sg_qa_vector_example
|
|
WHERE reference_status = 'APPROVED'
|
|
AND inspection_status = 'VERIFIED'
|
|
AND answer_sql IS NOT NULL
|
|
AND (source_type = 'POLICY_TEMPLATE'
|
|
OR NOT REGEXP_LIKE(answer_sql, '<[A-Z][A-Z0-9_]*>', 'i'))
|
|
AND (target_type = 'ANY' OR v_target_type = 'ANY' OR target_type = v_target_type)
|
|
) c
|
|
WHERE c.cosine_distance <= v_max_cosine_distance
|
|
) s
|
|
)
|
|
WHERE hybrid_score >= best_hybrid_score - v_hybrid_margin
|
|
ORDER BY hybrid_score DESC, cosine_distance, example_id
|
|
FETCH FIRST p_top_k ROWS ONLY;
|
|
RETURN v_results;
|
|
END;
|
|
/
|
|
|
|
COMMIT
|