feat(aihr): harden finance RAG retrieval and evidence gates

2026-08-03 22:34:18 +08:00
parent 99e8c509d5
commit 721839c847
22 changed files with 1306 additions and 153 deletions
@@ -0,0 +1,29 @@
+-- Generation-safe RAG candidate observability. Additive and idempotent on MySQL 8.
+
+CREATE TABLE IF NOT EXISTS `aihr_query_candidate_trace` (
+  `id` bigint NOT NULL AUTO_INCREMENT,
+  `tenant_id` varchar(20) NOT NULL,
+  `request_id` varchar(64) NOT NULL,
+  `fragment_id` bigint NOT NULL,
+  `candidate_rank` int DEFAULT NULL COMMENT 'Rank after rerank, null when rejected before rerank',
+  `channel` varchar(20) NOT NULL COMMENT 'KEYWORD/FULLTEXT/VECTOR',
+  `channel_rank` int NOT NULL,
+  `raw_score` decimal(18,10) DEFAULT NULL,
+  `fusion_score` decimal(18,10) DEFAULT NULL,
+  `rerank_score` decimal(18,10) DEFAULT NULL,
+  `rerank_model` varchar(100) DEFAULT NULL,
+  `rerank_applied` tinyint(1) NOT NULL DEFAULT 0,
+  `source_authority` varchar(30) DEFAULT NULL,
+  `source_kind` varchar(30) DEFAULT NULL,
+  `query_variant` varchar(30) NOT NULL COMMENT 'ORIGINAL/NORMALIZED/GLOSSARY/ALTERNATE_n',
+  `selected_for_evidence` tinyint(1) NOT NULL DEFAULT 0,
+  `decision_reason` varchar(64) NOT NULL,
+  `create_time` datetime NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  PRIMARY KEY (`id`),
+  UNIQUE KEY `uk_aihr_query_candidate_channel`
+    (`tenant_id`, `request_id`, `fragment_id`, `channel`, `query_variant`),
+  KEY `idx_aihr_query_candidate_request` (`tenant_id`, `request_id`, `candidate_rank`),
+  KEY `idx_aihr_query_candidate_reason` (`tenant_id`, `decision_reason`, `create_time`),
+  KEY `idx_aihr_query_candidate_fragment` (`tenant_id`, `fragment_id`, `create_time`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_0900_ai_ci
+  COMMENT='Pre-generation retrieval candidates, channel contributions, rerank and gate decisions';