АДРЕСНЫЙ РЕЖИМ - M2.3b тюнинг account-scope и диагностика стадий адресного рантайма

2026-03-29 20:57:55 +03:00
parent c82ebd70b7
commit 2bf16de4ea
498 changed files with 2619075 additions and 3 deletions
@@ -0,0 +1,28 @@
+# Run: 2026-03-29 Address Runtime Pivot Baseline
+
+Тип прогона: архитектурный baseline-аудит перед реализацией address-lane.
+
+## Цель
+
+Подтвердить фактическое состояние:
+
+1. текущий assistant runtime = snapshot-primary;
+2. live MCP = overlay и не для всех маршрутов;
+3. Postgres не является активным runtime-слоем в assistant path;
+4. требуется отдельный MCP-first lane для address-вопросов.
+
+## Состав артефактов
+
+- `run_summary.json`
+- `postgres_runtime_check.json`
+- `source_inventory.json`
+- `code_path_inventory.md`
+
+## Краткий результат
+
+Статус: `BASELINE_CONFIRMED`.
+
+Ключевой вывод:
+
+- технически проще и быстрее добавить отдельный `address_query` lane поверх текущего Stage 4, чем пытаться продолжать расширять snapshot-first path под новый класс задач.
+
@@ -0,0 +1,20 @@
+# Architecture Findings
+
+## Confirmed
+
+1. Current assistant execution is snapshot-primary, MCP-secondary.
+2. Large full snapshots exist, but assistant reads only narrow export samples.
+3. Existing claim-bound live path was designed for deep accounting families (VAT/RBP/FA), not address lookup UX.
+4. Canonical store DB layer is present (SQLite now, Postgres optional), but not part of active assistant runtime retrieval path.
+
+## Decision for next implementation wave
+
+1. Keep Stage 4 deep-chain flow unchanged.
+2. Add `address_query` lane with MCP-first source policy.
+3. Define explicit address intents, recipes, and response contract.
+4. Keep snapshot fallback explicit and limited (no silent substitution).
+
+## Baseline verdict
+
+`BASELINE_CONFIRMED_FOR_MCP_FIRST_ADDRESS_PIVOT`
+
@@ -0,0 +1,54 @@
+# Code Path Inventory (Address Pivot Baseline)
+
+## A. Assistant runtime data path
+
+1. `llm_normalizer/backend/src/services/assistantDataLayer.ts:2804`  
+   `AssistantDataLayer` uses `ARCH_EXPORT_2020_DIR` root by default.
+
+2. `llm_normalizer/backend/src/services/assistantDataLayer.ts:3379-3386`  
+   Runtime loads fixed sample files from `docs/ARCH/2020экспорт` (narrow package).
+
+3. `llm_normalizer/backend/src/services/assistantDataLayer.ts:2858-2863`  
+   MCP overlay is runtime-optional and route-limited.
+
+4. `llm_normalizer/backend/src/services/assistantDataLayer.ts:332`  
+   `buildLiveMcpCallPlan(...)` generates live calls.
+
+5. `llm_normalizer/backend/src/services/assistantDataLayer.ts:348,396,460,444`  
+   Claim-specific live plans exist for FA/VAT/RBP; else `claim_type: null`.
+
+## B. Routing discipline
+
+1. `llm_normalizer/backend/src/services/routeHintAdapter.ts:86-87`  
+   `exact_object_trace -> live_mcp_drilldown`.
+
+2. `llm_normalizer/backend/src/services/routeHintAdapter.ts:149-150`  
+   `canonical_fact_lookup -> store_canonical` (snapshot-oriented path).
+
+## C. MCP runtime config
+
+1. `llm_normalizer/backend/src/config.ts:90-100`  
+   MCP feature flag and defaults:
+   - `FEATURE_ASSISTANT_MCP_RUNTIME_V1` (default false),
+   - `ASSISTANT_MCP_PROXY_URL=http://127.0.0.1:6003`,
+   - `ASSISTANT_MCP_CHANNEL=default`.
+
+2. `llm_normalizer/package.json:11`  
+   `dev:all:mcp` enables MCP flags for local dev.
+
+## D. Canonical layer / DB
+
+1. `.env.example:18` and `README.md:89`  
+   Default canonical DB URL is SQLite.
+
+2. `canonical_layer/store.py:55,59`  
+   SQLAlchemy store supports non-SQLite URLs (including Postgres), but SQLite is default branch.
+
+3. `canonical_layer/app.py:61,82,96,128`  
+   Canonical API exposes read and refresh endpoints (`/documents`, `/postings`, `/counterparties/...`, `/refresh/run`).
+
+## E. Architectural implication
+
+Для Address V1 достаточно внедрить новый `address_query` lane в assistant runtime и сделать его MCP-first.
+Это не требует немедленной обязательной миграции в Postgres, но требует разведения route/intents и отдельного answer contract.
+
@@ -0,0 +1,20 @@
+{
+  "dataset": "docs/ARCH/2020_monthly_company_asof_full/_tmp_ndjson",
+  "total_files": 12,
+  "total_mb": 6732,
+  "files": [
+    { "name": "snapshot_2020-01.ndjson", "mb": 553.16 },
+    { "name": "snapshot_2020-02.ndjson", "mb": 554.27 },
+    { "name": "snapshot_2020-03.ndjson", "mb": 556.83 },
+    { "name": "snapshot_2020-04.ndjson", "mb": 557.63 },
+    { "name": "snapshot_2020-05.ndjson", "mb": 558.38 },
+    { "name": "snapshot_2020-06.ndjson", "mb": 561.38 },
+    { "name": "snapshot_2020-07.ndjson", "mb": 562.26 },
+    { "name": "snapshot_2020-08.ndjson", "mb": 563.23 },
+    { "name": "snapshot_2020-09.ndjson", "mb": 564.69 },
+    { "name": "snapshot_2020-10.ndjson", "mb": 565.01 },
+    { "name": "snapshot_2020-11.ndjson", "mb": 566.03 },
+    { "name": "snapshot_2020-12.ndjson", "mb": 569.14 }
+  ]
+}
+
@@ -0,0 +1,16 @@
+{
+  "check_date": "2026-03-29",
+  "postgres_process_detected": false,
+  "canonical_db_url_in_env_example": "sqlite:///X:/1C/NDC_1C/data/canonical_store.db",
+  "canonical_store_sqlite_file": {
+    "path": "X:/1C/NDC_1C/data/canonical_store.db",
+    "exists": true,
+    "size_bytes": 6594560
+  },
+  "assistant_runtime_direct_db_dependency_detected": false,
+  "notes": [
+    "Postgres is currently optional in canonical layer design, not mandatory in assistant runtime path.",
+    "Assistant runtime uses snapshot files and optional MCP overlay; it is not wired to canonical store reads."
+  ]
+}
+
@@ -0,0 +1,24 @@
+{
+  "run_id": "2026-03-29_Address_Runtime_Pivot_Baseline",
+  "date_utc": "2026-03-29",
+  "scope": "architecture_audit_for_mcp_first_address_queries",
+  "source_of_truth": [
+    "llm_normalizer/backend/src/services/assistantDataLayer.ts",
+    "llm_normalizer/backend/src/services/routeHintAdapter.ts",
+    "llm_normalizer/backend/src/config.ts",
+    "canonical_layer/store.py",
+    "canonical_layer/app.py",
+    ".env.example",
+    "README.md"
+  ],
+  "key_findings": [
+    "assistant runtime currently loads a narrow snapshot package (2020export) as primary source",
+    "live MCP is optional (feature-flagged) and route-limited overlay",
+    "current claim-bound live plan is focused on VAT/RBP/FA and does not define address-query intents",
+    "canonical layer supports SQLAlchemy db_url, but assistant runtime does not directly use canonical store",
+    "sqlite canonical store file exists, postgres process was not detected"
+  ],
+  "verdict": "MCP_INFRA_EXISTS_BUT_ADDRESS_RUNTIME_NOT_IMPLEMENTED",
+  "next_status_target": "ADDRESS_RUNTIME_V1_READY"
+}
+
@@ -0,0 +1,31 @@
+{
+  "snapshot_primary_runtime_package": {
+    "path": "X:/1C/NDC_1C/docs/ARCH/2020экспорт",
+    "files_total": 10,
+    "total_bytes": 1913578,
+    "total_mb": 1.82,
+    "files": [
+      "03_snapshot_fragment_problem_cases.json",
+      "04_samples_SpisanieSRaschetnogoScheta.json",
+      "05_samples_RealizaciyaTovarovUslug.json",
+      "06_samples_PostuplenieTovarovUslug.json",
+      "07_samples_DocumentJournals.json",
+      "08_samples_NDS_registers.json",
+      "09_samples_key_fields_Recorder_Ref_Supplier_Buyer_Responsible.json"
+    ]
+  },
+  "full_monthly_snapshot_corpus": {
+    "path": "X:/1C/NDC_1C/docs/ARCH/2020_monthly_company_asof_full/_tmp_ndjson",
+    "files_total": 12,
+    "total_bytes": 7059009054,
+    "total_mb": 6732,
+    "average_file_mb": 561
+  },
+  "live_mcp_channel_defaults": {
+    "proxy_url_default": "http://127.0.0.1:6003",
+    "channel_default": "default",
+    "feature_flag_default": false
+  },
+  "conclusion": "large_gap_between_available_snapshot_volume_and_runtime_loaded_snapshot_volume_confirmed"
+}
+