АДРЕСНЫЙ РЕЖИМ - M2.3e: стабилизация address runtime парсинг периодов, LLM pre-decompose fallback и приоритет intent для банковских операций

2026-04-01 19:35:25 +03:00
parent 4d59672576
commit 18e0f1364d
16 changed files with 1467 additions and 562 deletions
@@ -0,0 +1,59 @@
+# Address Query Runtime V1 - M2.3e Stability Hardening (Noise + Account Query Scope)
+
+Date: 2026-04-01 (updated)
+
+## Goal
+
+Finalize the stage before local LLM switch with production-safe stability:
+
+- robust noisy phrase handling;
+- deterministic month/account filtering;
+- safer LLM pre-decompose integration without semantic drift.
+
+## Implemented
+
+1. Noise hardening for counterparty extraction:
+- slang/noise tails (`плс`, `pls`, profanity markers) are ignored as anchors;
+- free-text heuristic for counterparty in document/bank requests without strict `по <anchor>`.
+
+2. Relaxed year extraction:
+- compact forms like `20 год` -> `2020-01-01..2020-12-31`;
+- standalone year mentions in noisy phrasing.
+
+3. Month period hardening for balance intents:
+- equivalent mapping for:
+  - `на май 2020`
+  - `на 2020.05`
+  - `на 2020 май`
+- all forms now converge to `period_from=2020-05-01`, `period_to=2020-05-31`, `as_of_date=2020-05-31`.
+
+4. Account query scope hardening in recipe:
+- account condition injected into movements query using account code fields;
+- removed fragile presentation-based filtering path for account subaccounts.
+
+5. LLM pre-decompose salvage path:
+- when strict normalized payload has no usable fragments, runtime now extracts a safe fragment from raw model JSON;
+- decomposition reason now can be `raw_fragment_applied` instead of always falling back to deterministic parsing.
+
+6. Intent routing hardening against LLM drift:
+- `bank_operations_by_counterparty` has higher priority than generic account fallback when party/bank signals are present;
+- strengthened Russian phrasing detection for `documents_forming_balance` (including participle forms like `формирующие остаток`).
+
+7. Test coverage expansion:
+- month parsing regressions (`на май 2020`, `на 2020.05`, `на 2020 май`);
+- bank-ops intent priority regression with injected account hints;
+- documents-forming-balance participle phrasing regression;
+- existing noisy-query and account-scope regressions retained.
+
+## Verification
+
+- `npx vitest tests/addressQueryRuntimeM23.test.ts` -> PASS (`46` tests)
+- `npx vitest tests/assistantAddressFollowupContext.test.ts` -> PASS (`1` test)
+- `npm run build` -> PASS
+
+## Notes
+
+- Architecture remains hybrid and stable: deterministic parser + guarded heuristics + optional LLM pre-decompose.
+- LLM pre-decompose is now resilient to schema drift in local model output.
+- No free-form query builder added.
+- No deep-analysis lane changes.
@@ -0,0 +1,29 @@
+{
+  "run_id": "2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope",
+  "comparison": {
+    "baseline_ref": "2026-04-01_Address_Query_Runtime_V1_M2_3D_Query_Variants_Expansion",
+    "current_ref": "workspace (after M2.3e stability update)"
+  },
+  "metrics": {
+    "address_m23_test_cases": {
+      "before": 39,
+      "after": 46,
+      "delta": 7
+    },
+    "address_m23_test_failures": {
+      "before": 0,
+      "after": 0,
+      "delta": 0
+    },
+    "llm_predecompose_signal": {
+      "before": "attempted_but_often_no_usable_fragment",
+      "after": "raw_fragment_salvage_available",
+      "delta": "stability_improved"
+    }
+  },
+  "notes": [
+    "delta includes month-format parsing regressions and intent-priority regressions",
+    "added fallback extraction from raw local LLM JSON when strict normalized fragments are empty",
+    "added regression checks for: 'на 2020.05', 'на 2020 май', bank ops with account hints"
+  ]
+}
@@ -0,0 +1,11 @@
+llm_normalizer/backend/src/services/addressFilterExtractor.ts
+llm_normalizer/backend/src/services/addressIntentResolver.ts
+llm_normalizer/backend/src/services/addressRecipeCatalog.ts
+llm_normalizer/backend/src/services/assistantService.ts
+llm_normalizer/backend/tests/addressQueryRuntimeM23.test.ts
+
+docs/ADDRESS/runs/2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope/README.md
+docs/ADDRESS/runs/2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope/run_summary.json
+docs/ADDRESS/runs/2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope/before_after_metrics.json
+docs/ADDRESS/runs/2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope/smoke_checks.md
+docs/ADDRESS/runs/2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope/changed_files.txt
@@ -0,0 +1,43 @@
+{
+  "run_id": "2026-04-01_Address_Query_Runtime_V1_M2_3E_Stability_Hardening_AccountQueryScope",
+  "date": "2026-04-01",
+  "stage": "Address Query Runtime V1",
+  "wave": "M2.3e",
+  "goal": "Stability hardening before local LLM switch: noisy phrase robustness + account/query period determinism + LLM pre-decompose resilience",
+  "status": "COMPLETED_UPDATED",
+  "scope": {
+    "new_intents": false,
+    "deep_analysis_changes": false,
+    "focus": [
+      "noisy phrase anchor stability",
+      "relaxed year extraction",
+      "month format normalization (month-year/year-month/year month)",
+      "account query scope in movements recipe",
+      "LLM pre-decompose raw-fragment salvage",
+      "intent priority hardening for bank ops vs account fallback",
+      "regression/unit coverage expansion"
+    ]
+  },
+  "checks": {
+    "test_command_primary": "npx vitest tests/addressQueryRuntimeM23.test.ts",
+    "test_command_secondary": "npx vitest tests/assistantAddressFollowupContext.test.ts",
+    "build_command": "npm run build",
+    "tests_passed": 47,
+    "tests_failed": 0
+  },
+  "guardrails": {
+    "false_factual_rate_target": 0,
+    "free_form_query_builder": "not_added",
+    "whitelist_recipe_policy": "unchanged"
+  },
+  "key_changes": {
+    "noise_tokens_not_used_as_anchor": true,
+    "free_text_counterparty_heuristic": true,
+    "compact_year_support": true,
+    "month_year_variants_supported": true,
+    "movements_query_account_condition_injected": true,
+    "llm_predecompose_raw_salvage": true,
+    "intent_priority_bank_ops_over_account_fallback": true,
+    "documents_forming_balance_participle_detection": true
+  }
+}
@@ -0,0 +1,23 @@
+# Smoke Checks
+
+## Backend tests
+
+- Command: `npx vitest tests/addressQueryRuntimeM23.test.ts`
+- Result: PASS
+- Details: `1 passed file`, `46 passed tests`, `0 failed`
+
+- Command: `npx vitest tests/assistantAddressFollowupContext.test.ts`
+- Result: PASS
+- Details: `1 passed file`, `1 passed test`, `0 failed`
+
+## Build
+
+- Command: `npm run build`
+- Result: PASS
+- Details: TypeScript build completed without errors.
+
+## Manual spot-check (address lane)
+
+- File: `docs/ADDRESS/1.txt`
+- Result: month-variant account questions produce consistent May-2020 window.
+- Result: last bank-operations question now covered by intent-priority hardening (expected recipe: bank ops by counterparty).