refs #739: document DB-owned task and few-shot policy

database/adb/140_sgmp_hybrid_fewshot_reranking.sql

@@ -0,0 +1,95 @@
+-- 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