egullickson/motovaultpro
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Replace NHTSA VIN decode with Google Gemini via OCR service #223

New Issue
Closed
opened 2026-02-19 03:11:02 +00:00 by egullickson · 5 comments
egullickson commented 2026-02-19 03:11:02 +00:00
Owner
Copy Link

Summary

Replace the NHTSA vPIC API integration for VIN string decoding with Google Gemini, routed through the Python OCR microservice (mvp-ocr) to match the existing receipt/manual extraction architecture.

Current State

  • POST /api/vehicles/decode-vin calls NHTSA vPIC API directly from the TypeScript backend (vehicles/external/nhtsa/)
  • Returns vehicle details: year, make, model, trim, engine, transmission, bodyType, driveType, fuelType
  • Results cached in vin_cache database table (1-year TTL)
  • Pro/Enterprise tier gated (vehicle.vinDecode)
  • OCR image-based VIN extraction (POST /api/ocr/extract/vin) is separate and unchanged by this issue

Expected Behavior

  • VIN string decode uses Google Gemini (same workflow as receipt extraction)
  • Flow: Backend proxy -> Python OCR service -> Gemini -> structured vehicle data
  • Same endpoint and response shape for POST /api/vehicles/decode-vin
  • Keep vin_cache table for caching Gemini decode results
  • Remove all NHTSA vPIC API code and references from the codebase

Scope

Remove

  • vehicles/external/nhtsa/ directory (nhtsa.client.ts, nhtsa.types.ts, index.ts)
  • All NHTSA references in vehicles service, tests, and documentation
  • NHTSA references in platform feature documentation (planned fallback strategy)

Add/Modify

  • New VIN decode endpoint in Python OCR service using Gemini
  • Backend OCR client method for VIN string decode
  • Update vehicles service to route VIN decode through OCR proxy
  • Update tests for new flow
  • Update documentation (platform README, vehicles README, OCR README)

Acceptance Criteria

  • VIN decode returns same response shape (year, make, model, trim, etc. with confidence)
  • VIN decode routes through mvp-ocr Python service using Gemini
  • vin_cache table still caches decode results
  • Zero NHTSA references remain in the codebase
  • Existing OCR image-based VIN extraction (/api/ocr/extract/vin) unaffected
  • All tests pass
  • Tier gating unchanged (Pro/Enterprise)
## Summary Replace the NHTSA vPIC API integration for VIN string decoding with Google Gemini, routed through the Python OCR microservice (mvp-ocr) to match the existing receipt/manual extraction architecture. ## Current State - `POST /api/vehicles/decode-vin` calls NHTSA vPIC API directly from the TypeScript backend (`vehicles/external/nhtsa/`) - Returns vehicle details: year, make, model, trim, engine, transmission, bodyType, driveType, fuelType - Results cached in `vin_cache` database table (1-year TTL) - Pro/Enterprise tier gated (`vehicle.vinDecode`) - OCR image-based VIN extraction (`POST /api/ocr/extract/vin`) is separate and unchanged by this issue ## Expected Behavior - VIN string decode uses Google Gemini (same workflow as receipt extraction) - Flow: Backend proxy -> Python OCR service -> Gemini -> structured vehicle data - Same endpoint and response shape for `POST /api/vehicles/decode-vin` - Keep `vin_cache` table for caching Gemini decode results - Remove all NHTSA vPIC API code and references from the codebase ## Scope ### Remove - `vehicles/external/nhtsa/` directory (nhtsa.client.ts, nhtsa.types.ts, index.ts) - All NHTSA references in vehicles service, tests, and documentation - NHTSA references in platform feature documentation (planned fallback strategy) ### Add/Modify - New VIN decode endpoint in Python OCR service using Gemini - Backend OCR client method for VIN string decode - Update vehicles service to route VIN decode through OCR proxy - Update tests for new flow - Update documentation (platform README, vehicles README, OCR README) ## Acceptance Criteria - [ ] VIN decode returns same response shape (year, make, model, trim, etc. with confidence) - [ ] VIN decode routes through mvp-ocr Python service using Gemini - [ ] `vin_cache` table still caches decode results - [ ] Zero NHTSA references remain in the codebase - [ ] Existing OCR image-based VIN extraction (`/api/ocr/extract/vin`) unaffected - [ ] All tests pass - [ ] Tier gating unchanged (Pro/Enterprise)
egullickson added the
status
backlog
type
feature
 labels 2026-02-19 03:11:06 +00:00
egullickson added
status
in-progress
 and removed
status
backlog
 labels 2026-02-19 03:13:34 +00:00
egullickson commented 2026-02-19 03:20:43 +00:00
Author
Owner
Copy Link

Plan: Replace NHTSA VIN decode with Google Gemini via OCR service

Phase: Planning | Agent: Planner | Status: AWAITING_REVIEW

Architecture Decision

Decision: Add decode_vin() method to existing GeminiEngine class (not a new standalone module).
Rationale: GeminiEngine already has lazy Vertex AI init, WIF credentials, and model setup. VIN decode is the same pattern (text-to-structured-data via Gemini). Duplicating init in a separate class adds complexity without SRP benefit at 2 methods. Decision Critic verdict: STAND (all claims verified, risks minor).

Current Flow

Frontend -> POST /api/vehicles/decode-vin -> VehiclesController
  -> NHTSAClient.decodeVin() -> NHTSA vPIC API (external)
  -> NHTSAClient cache (vin_cache table)
  -> VehiclesService.mapNHTSAResponse() -> dropdown matching
  -> DecodedVehicleData response

New Flow

Frontend -> POST /api/vehicles/decode-vin -> VehiclesController
  -> vin_cache check (moved to controller/service)
  -> OcrClient.decodeVin() -> POST /decode/vin on mvp-ocr
    -> GeminiEngine.decode_vin() -> Gemini 2.5 Flash (Vertex AI)
  -> cache Gemini response in vin_cache
  -> VehiclesService.mapVinDecodeResponse() -> dropdown matching
  -> DecodedVehicleData response (nhtsaValue renamed to sourceValue)

Key Changes

  • nhtsaValue renamed to sourceValue in MatchedField<T> (backend + frontend)
  • VIN validation regex moves from NHTSAClient to vehicles controller
  • Cache logic moves from NHTSAClient to vehicles service
  • mapNHTSAResponse() becomes mapVinDecodeResponse() - simpler since Gemini returns structured fields directly (no NHTSA array parsing)
  • Migration truncates vin_cache to clear NHTSA-format entries

Milestones

M1: OCR Python endpoint (#224)

Files: ocr/app/engines/gemini_engine.py, ocr/app/routers/extract.py, ocr/app/models/*.py

  • Add decode_vin(vin: str) to GeminiEngine with VIN-specific prompt and JSON response schema
  • Add VinDecodeResponse Pydantic model
  • Add POST /decode/vin route accepting {"vin": "..."} JSON body
  • VIN format validation (17 chars, regex ^[A-HJ-NPR-Z0-9]{17}$)
  • Returns: year, make, model, trimLevel, bodyType, driveType, fuelType, engine, transmission, confidence

M2: Backend types and OCR client (#225)

Files: backend/src/features/ocr/domain/ocr.types.ts, backend/src/features/ocr/external/ocr-client.ts, backend/src/features/vehicles/domain/vehicles.types.ts (new, extracted from nhtsa.types.ts)

  • Add VinDecodeResponse type matching OCR service response
  • Add decodeVin(vin: string) to OcrClient (JSON POST, not multipart)
  • Extract MatchedField<T>, DecodedVehicleData, DecodeVinRequest, VinDecodeError from nhtsa.types.ts into vehicles types
  • Rename nhtsaValue to sourceValue in MatchedField<T>

M3: Rewire vehicles controller (#226)

Files: backend/src/features/vehicles/api/vehicles.controller.ts, backend/src/features/vehicles/domain/vehicles.service.ts, backend/src/features/vehicles/api/vehicles.routes.ts, backend/src/features/vehicles/migrations/007_truncate_vin_cache.sql

  • Replace NHTSAClient import with OcrClient in controller
  • Move VIN validation into controller
  • Move cache read/write (getCached/saveToCache) into vehicles service
  • Refactor mapNHTSAResponse() to mapVinDecodeResponse() - takes flat Gemini fields, does dropdown matching
  • Add 007_truncate_vin_cache.sql migration
  • Update error handling references

M4: Remove NHTSA code and update docs (#227)

Files: Delete backend/src/features/vehicles/external/nhtsa/ (3 files), modify backend/src/features/platform/models/responses.ts, backend/src/features/platform/README.md, 5 CLAUDE.md files

  • Delete entire vehicles/external/nhtsa/ directory
  • Remove VPICVariable, VPICResponse from platform models
  • Update NHTSA references in platform README
  • Update all affected CLAUDE.md navigation indexes

M5: Frontend and tests (#228)

Files: frontend/src/features/vehicles/types/vehicles.types.ts, frontend/src/features/vehicles/components/VinOcrReviewModal.tsx, guide page sections, test files

  • Rename nhtsaValue to sourceValue in frontend types
  • Update VinOcrReviewModal references
  • Update NHTSA comments in guide sections
  • Update integration/unit tests for new OCR-based flow
  • Verify existing OCR image VIN extraction unaffected

Risk Assessment

  • Cache migration: TRUNCATE is safe - cache is rebuild-on-miss, no data loss
  • Gemini accuracy: VIN decode is well-understood domain, Gemini handles it reliably
  • Response shape: Same structure preserved, only field rename (nhtsaValue -> sourceValue)
  • Existing OCR VIN extraction: Image-based /api/ocr/extract/vin endpoint is completely separate and unaffected

Files Summary

Action Count Details
DELETE 3 vehicles/external/nhtsa/ (index.ts, nhtsa.client.ts, nhtsa.types.ts)
NEW 2 vehicles.types.ts (extracted), 007_truncate_vin_cache.sql
MODIFY ~15 OCR engine, router, models, OcrClient, controller, service, routes, platform models/docs, CLAUDE.md files, frontend types/components/tests

Verdict: AWAITING_REVIEW | Next: Plan review cycle (QR plan-completeness)

## Plan: Replace NHTSA VIN decode with Google Gemini via OCR service **Phase**: Planning | **Agent**: Planner | **Status**: AWAITING_REVIEW ### Architecture Decision **Decision**: Add `decode_vin()` method to existing `GeminiEngine` class (not a new standalone module). **Rationale**: GeminiEngine already has lazy Vertex AI init, WIF credentials, and model setup. VIN decode is the same pattern (text-to-structured-data via Gemini). Duplicating init in a separate class adds complexity without SRP benefit at 2 methods. Decision Critic verdict: STAND (all claims verified, risks minor). ### Current Flow ``` Frontend -> POST /api/vehicles/decode-vin -> VehiclesController -> NHTSAClient.decodeVin() -> NHTSA vPIC API (external) -> NHTSAClient cache (vin_cache table) -> VehiclesService.mapNHTSAResponse() -> dropdown matching -> DecodedVehicleData response ``` ### New Flow ``` Frontend -> POST /api/vehicles/decode-vin -> VehiclesController -> vin_cache check (moved to controller/service) -> OcrClient.decodeVin() -> POST /decode/vin on mvp-ocr -> GeminiEngine.decode_vin() -> Gemini 2.5 Flash (Vertex AI) -> cache Gemini response in vin_cache -> VehiclesService.mapVinDecodeResponse() -> dropdown matching -> DecodedVehicleData response (nhtsaValue renamed to sourceValue) ``` ### Key Changes - `nhtsaValue` renamed to `sourceValue` in `MatchedField<T>` (backend + frontend) - VIN validation regex moves from NHTSAClient to vehicles controller - Cache logic moves from NHTSAClient to vehicles service - `mapNHTSAResponse()` becomes `mapVinDecodeResponse()` - simpler since Gemini returns structured fields directly (no NHTSA array parsing) - Migration truncates `vin_cache` to clear NHTSA-format entries ### Milestones #### M1: OCR Python endpoint (#224) **Files**: `ocr/app/engines/gemini_engine.py`, `ocr/app/routers/extract.py`, `ocr/app/models/*.py` - Add `decode_vin(vin: str)` to GeminiEngine with VIN-specific prompt and JSON response schema - Add `VinDecodeResponse` Pydantic model - Add `POST /decode/vin` route accepting `{"vin": "..."}` JSON body - VIN format validation (17 chars, regex `^[A-HJ-NPR-Z0-9]{17}$`) - Returns: year, make, model, trimLevel, bodyType, driveType, fuelType, engine, transmission, confidence #### M2: Backend types and OCR client (#225) **Files**: `backend/src/features/ocr/domain/ocr.types.ts`, `backend/src/features/ocr/external/ocr-client.ts`, `backend/src/features/vehicles/domain/vehicles.types.ts` (new, extracted from nhtsa.types.ts) - Add `VinDecodeResponse` type matching OCR service response - Add `decodeVin(vin: string)` to `OcrClient` (JSON POST, not multipart) - Extract `MatchedField<T>`, `DecodedVehicleData`, `DecodeVinRequest`, `VinDecodeError` from nhtsa.types.ts into vehicles types - Rename `nhtsaValue` to `sourceValue` in `MatchedField<T>` #### M3: Rewire vehicles controller (#226) **Files**: `backend/src/features/vehicles/api/vehicles.controller.ts`, `backend/src/features/vehicles/domain/vehicles.service.ts`, `backend/src/features/vehicles/api/vehicles.routes.ts`, `backend/src/features/vehicles/migrations/007_truncate_vin_cache.sql` - Replace NHTSAClient import with OcrClient in controller - Move VIN validation into controller - Move cache read/write (getCached/saveToCache) into vehicles service - Refactor `mapNHTSAResponse()` to `mapVinDecodeResponse()` - takes flat Gemini fields, does dropdown matching - Add `007_truncate_vin_cache.sql` migration - Update error handling references #### M4: Remove NHTSA code and update docs (#227) **Files**: Delete `backend/src/features/vehicles/external/nhtsa/` (3 files), modify `backend/src/features/platform/models/responses.ts`, `backend/src/features/platform/README.md`, 5 CLAUDE.md files - Delete entire `vehicles/external/nhtsa/` directory - Remove `VPICVariable`, `VPICResponse` from platform models - Update NHTSA references in platform README - Update all affected CLAUDE.md navigation indexes #### M5: Frontend and tests (#228) **Files**: `frontend/src/features/vehicles/types/vehicles.types.ts`, `frontend/src/features/vehicles/components/VinOcrReviewModal.tsx`, guide page sections, test files - Rename `nhtsaValue` to `sourceValue` in frontend types - Update VinOcrReviewModal references - Update NHTSA comments in guide sections - Update integration/unit tests for new OCR-based flow - Verify existing OCR image VIN extraction unaffected ### Risk Assessment - **Cache migration**: TRUNCATE is safe - cache is rebuild-on-miss, no data loss - **Gemini accuracy**: VIN decode is well-understood domain, Gemini handles it reliably - **Response shape**: Same structure preserved, only field rename (`nhtsaValue` -> `sourceValue`) - **Existing OCR VIN extraction**: Image-based `/api/ocr/extract/vin` endpoint is completely separate and unaffected ### Files Summary | Action | Count | Details | |--------|-------|---------| | DELETE | 3 | `vehicles/external/nhtsa/` (index.ts, nhtsa.client.ts, nhtsa.types.ts) | | NEW | 2 | vehicles.types.ts (extracted), 007_truncate_vin_cache.sql | | MODIFY | ~15 | OCR engine, router, models, OcrClient, controller, service, routes, platform models/docs, CLAUDE.md files, frontend types/components/tests | *Verdict*: AWAITING_REVIEW | *Next*: Plan review cycle (QR plan-completeness)
egullickson commented 2026-02-19 03:25:27 +00:00
Author
Owner
Copy Link

Plan v2: Replace NHTSA VIN decode with Google Gemini via OCR service

Phase: Planning | Agent: Planner | Status: AWAITING_REVIEW

Revised based on QR plan-completeness FAIL and TW plan-scrub FAIL. Addresses: response schema, sequencing conflict, incomplete file lists, cache type disposition, prompt design, route registration.

Architecture Decision

Decision: Add decode_vin() method to existing GeminiEngine class.
Rationale: Reuses lazy Vertex AI init and WIF credentials. Decision Critic verdict: STAND.

New Flow

Frontend -> POST /api/vehicles/decode-vin -> VehiclesController
  -> vin_cache check (in VehiclesService)
  -> OcrClient.decodeVin(vin) -> POST /decode/vin on mvp-ocr
    -> GeminiEngine.decode_vin(vin) -> Gemini 2.5 Flash
  -> VehiclesService.saveVinCache(vin, response)
  -> VehiclesService.mapVinDecodeResponse(response) -> dropdown matching
  -> DecodedVehicleData response (sourceValue field)

OCR VIN Decode Response Schema (drives M1 and M2)

{
  "success": true,
  "vin": "1HGBH41JXMN109186",
  "year": 2020,
  "make": "Toyota",
  "model": "Camry",
  "trimLevel": "SE",
  "bodyType": "Sedan 4-Door",
  "driveType": "Front-Wheel Drive",
  "fuelType": "Gasoline",
  "engine": "2.5L 4-Cylinder",
  "transmission": "8-Speed Automatic",
  "confidence": 0.95,
  "processingTimeMs": 1200,
  "error": null
}

Gemini Prompt Design (M1)

Given the VIN (Vehicle Identification Number) below, decode it and return the vehicle specifications.

VIN: {vin}

Return the vehicle's year, make, model, trim level, body type, drive type, fuel type, engine description, and transmission type. If a field cannot be determined from the VIN, return null for that field. Return a confidence score (0.0-1.0) indicating overall decode reliability.

Response schema enforced via GenerationConfig(response_mime_type="application/json", response_schema=_VIN_DECODE_SCHEMA).

Milestones

M1: OCR Python VIN decode endpoint (#224)

Files:

  • ocr/app/engines/gemini_engine.py - add decode_vin(vin: str) method with _VIN_DECODE_PROMPT and _VIN_DECODE_SCHEMA
  • ocr/app/routers/decode.py (NEW) - new router with prefix /decode, endpoint POST /vin
  • ocr/app/routers/__init__.py - export decode_router
  • ocr/app/main.py - register decode_router
  • ocr/app/models/__init__.py (or models file) - add VinDecodeResponse Pydantic model
  • ocr/tests/test_vin_decode.py (NEW) - unit test for /decode/vin route

Implementation details:

  • decode_vin() calls model.generate_content([prompt_text], generation_config=vin_config) - text-only input, no Part.from_data
  • Separate _generation_config per method (maintenance schema vs VIN schema), stored as instance vars on first use
  • VIN format validation in the route handler: regex ^[A-HJ-NPR-Z0-9]{17}$, return 400 on failure
  • Route returns VinDecodeResponse Pydantic model matching the schema above

M2: Backend OCR client VIN decode method (#225)

Files:

  • backend/src/features/ocr/domain/ocr.types.ts - add VinDecodeResponse type
  • backend/src/features/ocr/external/ocr-client.ts - add decodeVin(vin: string) method

Implementation details:

  • OcrClient.decodeVin() uses JSON POST (not multipart): fetch(url, { method: 'POST', headers: {'Content-Type': 'application/json'}, body: JSON.stringify({vin}) })
  • Returns VinDecodeResponse matching the Python schema above
  • This is a departure from existing OcrClient methods (which use multipart) - this method sends JSON because VIN decode has no file upload

Not in this milestone: Type extraction and nhtsaValue rename moved to M3 to avoid sequencing breakage.

M3: Rewire vehicles controller and service (#226)

Files:

  • backend/src/features/vehicles/api/vehicles.controller.ts - replace NHTSAClient with OcrClient, move VIN regex validation here
  • backend/src/features/vehicles/domain/vehicles.service.ts - add getVinCached(), saveVinCache(), rename mapNHTSAResponse() to mapVinDecodeResponse(response: VinDecodeResponse), rename nhtsaValue to sourceValue in matchField() and all MatchedField construction
  • backend/src/features/vehicles/domain/vehicles.types.ts (existing file) - add MatchedField<T>, MatchConfidence, DecodedVehicleData, DecodeVinRequest, VinDecodeError, VinCacheEntry (all extracted from nhtsa.types.ts with nhtsaValue -> sourceValue rename and rawData type changed to VinDecodeResponse)
  • backend/src/features/vehicles/api/vehicles.routes.ts - update comment from "NHTSA vPIC API" to "OCR VIN decode"
  • backend/src/features/vehicles/migrations/007_truncate_vin_cache.sql - TRUNCATE TABLE vin_cache;

Implementation details:

  • Controller decodeVin() method flow:
    1. Validate VIN format (regex ^[A-HJ-NPR-Z0-9]{17}$), return 400 if invalid
    2. Call vehiclesService.getVinCached(vin) - if hit, return cached DecodedVehicleData
    3. Call ocrClient.decodeVin(vin) to get VinDecodeResponse
    4. Call vehiclesService.saveVinCache(vin, response) to cache
    5. Call vehiclesService.mapVinDecodeResponse(response) for dropdown matching
    6. Return DecodedVehicleData
  • mapVinDecodeResponse(response: VinDecodeResponse): Promise<DecodedVehicleData>:
    • Reads flat fields directly: response.year, response.make, etc. (no NHTSAClient.extractValue calls)
    • matchField() logic unchanged (fuzzy matching against dropdowns)
    • All MatchedField objects use sourceValue instead of nhtsaValue
  • VinCacheEntry.rawData type changes from NHTSADecodeResponse to VinDecodeResponse
  • saveVinCache() stores Gemini response as raw_data JSONB
  • getVinCached() returns cached VinDecodeResponse from raw_data, passes through mapVinDecodeResponse() if cache hit (so dropdown matching always runs fresh)
  • Error handling: replace "NHTSA API request timed out" with "VIN decode service timed out", replace "NHTSA" error check with "OCR" error check

M4: Remove NHTSA code and update docs (#227)

Files to DELETE:

  • backend/src/features/vehicles/external/nhtsa/index.ts
  • backend/src/features/vehicles/external/nhtsa/nhtsa.client.ts
  • backend/src/features/vehicles/external/nhtsa/nhtsa.types.ts

Files to MODIFY:

  • backend/src/features/platform/models/responses.ts - remove VPICVariable and VPICResponse interfaces (lines 65-79). Verified: no other files import these types.
  • backend/src/features/platform/README.md - remove NHTSA vPIC fallback strategy section, update VIN decode references to point to OCR service
  • backend/src/features/vehicles/README.md - update VIN decode section from NHTSA to Gemini/OCR
  • backend/src/features/vehicles/CLAUDE.md - change external/ description from "External service integrations (NHTSA)" to "External service integrations"
  • backend/src/features/vehicles/external/CLAUDE.md - remove nhtsa/ directory entry
  • ocr/app/engines/CLAUDE.md - add VIN decode to GeminiEngine description
  • ocr/CLAUDE.md - add "Gemini VIN decode" to service description
  • ocr/app/CLAUDE.md - add decode router to subdirectories
  • backend/src/features/ocr/CLAUDE.md - add decodeVin to OcrClient method list

M5: Frontend updates and tests (#228)

Frontend files:

  • frontend/src/features/vehicles/types/vehicles.types.ts - rename nhtsaValue to sourceValue in MatchedField<T>, update DecodedVehicleData comment from "NHTSA vPIC API" to "VIN decode"
  • frontend/src/features/vehicles/components/VinOcrReviewModal.tsx - rename nhtsaValue references (lines 114-118) to sourceValue, update nhtsaRefs variable name to sourceRefs
  • frontend/src/features/vehicles/hooks/useVinOcr.ts - update step 2 label from "Decode VIN using NHTSA" to "Decode VIN", update error messages removing NHTSA
  • frontend/src/features/vehicles/api/vehicles.api.ts - update JSDoc from "NHTSA vPIC API" to "VIN decode service"
  • frontend/src/features/vehicles/components/VehicleForm.tsx - update NHTSA comment (line 510)
  • frontend/src/pages/GuidePage/sections/VehiclesSection.tsx - change "NHTSA database" to "vehicle database"
  • frontend/src/pages/GuidePage/sections/SubscriptionSection.tsx - change "NHTSA database" to "vehicle database"

Test files:

  • backend/src/features/vehicles/tests/integration/vehicles.integration.test.ts - update vin_cache references if needed
  • Add unit tests for VehiclesService.mapVinDecodeResponse() mocking OcrClient
  • Add unit test for OcrClient.decodeVin() method

Regression verification:

  • Confirm POST /api/ocr/extract/vin (image-based VIN extraction) still works - run existing OCR tests, verify PaddleOCR path untouched

Milestone Dependencies

M1 (OCR endpoint) -> M2 (backend client) -> M3 (rewire controller + types + rename)
                                                -> M4 (delete NHTSA, must happen after M3 moves all code)
                                                -> M5 (frontend + tests, after M3 stabilizes API)

Files Summary

Action Count Details
DELETE 3 vehicles/external/nhtsa/ directory
NEW 3 ocr/app/routers/decode.py, ocr/tests/test_vin_decode.py, 007_truncate_vin_cache.sql
MODIFY ~20 OCR engine/models/main/routers, OcrClient/types, vehicles controller/service/types/routes, platform models/README, 8 CLAUDE.md files, 7 frontend files, test files

Verdict: AWAITING_REVIEW | Next: QR plan-code, QR plan-docs

## Plan v2: Replace NHTSA VIN decode with Google Gemini via OCR service **Phase**: Planning | **Agent**: Planner | **Status**: AWAITING_REVIEW *Revised based on QR plan-completeness FAIL and TW plan-scrub FAIL. Addresses: response schema, sequencing conflict, incomplete file lists, cache type disposition, prompt design, route registration.* ### Architecture Decision **Decision**: Add `decode_vin()` method to existing `GeminiEngine` class. **Rationale**: Reuses lazy Vertex AI init and WIF credentials. Decision Critic verdict: STAND. ### New Flow ``` Frontend -> POST /api/vehicles/decode-vin -> VehiclesController -> vin_cache check (in VehiclesService) -> OcrClient.decodeVin(vin) -> POST /decode/vin on mvp-ocr -> GeminiEngine.decode_vin(vin) -> Gemini 2.5 Flash -> VehiclesService.saveVinCache(vin, response) -> VehiclesService.mapVinDecodeResponse(response) -> dropdown matching -> DecodedVehicleData response (sourceValue field) ``` ### OCR VIN Decode Response Schema (drives M1 and M2) ```json { "success": true, "vin": "1HGBH41JXMN109186", "year": 2020, "make": "Toyota", "model": "Camry", "trimLevel": "SE", "bodyType": "Sedan 4-Door", "driveType": "Front-Wheel Drive", "fuelType": "Gasoline", "engine": "2.5L 4-Cylinder", "transmission": "8-Speed Automatic", "confidence": 0.95, "processingTimeMs": 1200, "error": null } ``` ### Gemini Prompt Design (M1) ``` Given the VIN (Vehicle Identification Number) below, decode it and return the vehicle specifications. VIN: {vin} Return the vehicle's year, make, model, trim level, body type, drive type, fuel type, engine description, and transmission type. If a field cannot be determined from the VIN, return null for that field. Return a confidence score (0.0-1.0) indicating overall decode reliability. ``` Response schema enforced via `GenerationConfig(response_mime_type="application/json", response_schema=_VIN_DECODE_SCHEMA)`. --- ### Milestones #### M1: OCR Python VIN decode endpoint (#224) **Files**: - `ocr/app/engines/gemini_engine.py` - add `decode_vin(vin: str)` method with `_VIN_DECODE_PROMPT` and `_VIN_DECODE_SCHEMA` - `ocr/app/routers/decode.py` (NEW) - new router with prefix `/decode`, endpoint `POST /vin` - `ocr/app/routers/__init__.py` - export `decode_router` - `ocr/app/main.py` - register `decode_router` - `ocr/app/models/__init__.py` (or models file) - add `VinDecodeResponse` Pydantic model - `ocr/tests/test_vin_decode.py` (NEW) - unit test for `/decode/vin` route **Implementation details**: - `decode_vin()` calls `model.generate_content([prompt_text], generation_config=vin_config)` - text-only input, no `Part.from_data` - Separate `_generation_config` per method (maintenance schema vs VIN schema), stored as instance vars on first use - VIN format validation in the route handler: regex `^[A-HJ-NPR-Z0-9]{17}$`, return 400 on failure - Route returns `VinDecodeResponse` Pydantic model matching the schema above #### M2: Backend OCR client VIN decode method (#225) **Files**: - `backend/src/features/ocr/domain/ocr.types.ts` - add `VinDecodeResponse` type - `backend/src/features/ocr/external/ocr-client.ts` - add `decodeVin(vin: string)` method **Implementation details**: - `OcrClient.decodeVin()` uses JSON POST (not multipart): `fetch(url, { method: 'POST', headers: {'Content-Type': 'application/json'}, body: JSON.stringify({vin}) })` - Returns `VinDecodeResponse` matching the Python schema above - This is a departure from existing OcrClient methods (which use multipart) - this method sends JSON because VIN decode has no file upload **Not in this milestone**: Type extraction and nhtsaValue rename moved to M3 to avoid sequencing breakage. #### M3: Rewire vehicles controller and service (#226) **Files**: - `backend/src/features/vehicles/api/vehicles.controller.ts` - replace NHTSAClient with OcrClient, move VIN regex validation here - `backend/src/features/vehicles/domain/vehicles.service.ts` - add `getVinCached()`, `saveVinCache()`, rename `mapNHTSAResponse()` to `mapVinDecodeResponse(response: VinDecodeResponse)`, rename `nhtsaValue` to `sourceValue` in `matchField()` and all MatchedField construction - `backend/src/features/vehicles/domain/vehicles.types.ts` (existing file) - add `MatchedField<T>`, `MatchConfidence`, `DecodedVehicleData`, `DecodeVinRequest`, `VinDecodeError`, `VinCacheEntry` (all extracted from nhtsa.types.ts with `nhtsaValue` -> `sourceValue` rename and `rawData` type changed to `VinDecodeResponse`) - `backend/src/features/vehicles/api/vehicles.routes.ts` - update comment from "NHTSA vPIC API" to "OCR VIN decode" - `backend/src/features/vehicles/migrations/007_truncate_vin_cache.sql` - `TRUNCATE TABLE vin_cache;` **Implementation details**: - Controller `decodeVin()` method flow: 1. Validate VIN format (regex `^[A-HJ-NPR-Z0-9]{17}$`), return 400 if invalid 2. Call `vehiclesService.getVinCached(vin)` - if hit, return cached `DecodedVehicleData` 3. Call `ocrClient.decodeVin(vin)` to get `VinDecodeResponse` 4. Call `vehiclesService.saveVinCache(vin, response)` to cache 5. Call `vehiclesService.mapVinDecodeResponse(response)` for dropdown matching 6. Return `DecodedVehicleData` - `mapVinDecodeResponse(response: VinDecodeResponse): Promise<DecodedVehicleData>`: - Reads flat fields directly: `response.year`, `response.make`, etc. (no NHTSAClient.extractValue calls) - `matchField()` logic unchanged (fuzzy matching against dropdowns) - All `MatchedField` objects use `sourceValue` instead of `nhtsaValue` - `VinCacheEntry.rawData` type changes from `NHTSADecodeResponse` to `VinDecodeResponse` - `saveVinCache()` stores Gemini response as `raw_data` JSONB - `getVinCached()` returns cached `VinDecodeResponse` from `raw_data`, passes through `mapVinDecodeResponse()` if cache hit (so dropdown matching always runs fresh) - Error handling: replace "NHTSA API request timed out" with "VIN decode service timed out", replace "NHTSA" error check with "OCR" error check #### M4: Remove NHTSA code and update docs (#227) **Files to DELETE**: - `backend/src/features/vehicles/external/nhtsa/index.ts` - `backend/src/features/vehicles/external/nhtsa/nhtsa.client.ts` - `backend/src/features/vehicles/external/nhtsa/nhtsa.types.ts` **Files to MODIFY**: - `backend/src/features/platform/models/responses.ts` - remove `VPICVariable` and `VPICResponse` interfaces (lines 65-79). Verified: no other files import these types. - `backend/src/features/platform/README.md` - remove NHTSA vPIC fallback strategy section, update VIN decode references to point to OCR service - `backend/src/features/vehicles/README.md` - update VIN decode section from NHTSA to Gemini/OCR - `backend/src/features/vehicles/CLAUDE.md` - change `external/` description from "External service integrations (NHTSA)" to "External service integrations" - `backend/src/features/vehicles/external/CLAUDE.md` - remove `nhtsa/` directory entry - `ocr/app/engines/CLAUDE.md` - add VIN decode to GeminiEngine description - `ocr/CLAUDE.md` - add "Gemini VIN decode" to service description - `ocr/app/CLAUDE.md` - add decode router to subdirectories - `backend/src/features/ocr/CLAUDE.md` - add `decodeVin` to OcrClient method list #### M5: Frontend updates and tests (#228) **Frontend files**: - `frontend/src/features/vehicles/types/vehicles.types.ts` - rename `nhtsaValue` to `sourceValue` in `MatchedField<T>`, update `DecodedVehicleData` comment from "NHTSA vPIC API" to "VIN decode" - `frontend/src/features/vehicles/components/VinOcrReviewModal.tsx` - rename `nhtsaValue` references (lines 114-118) to `sourceValue`, update `nhtsaRefs` variable name to `sourceRefs` - `frontend/src/features/vehicles/hooks/useVinOcr.ts` - update step 2 label from "Decode VIN using NHTSA" to "Decode VIN", update error messages removing NHTSA - `frontend/src/features/vehicles/api/vehicles.api.ts` - update JSDoc from "NHTSA vPIC API" to "VIN decode service" - `frontend/src/features/vehicles/components/VehicleForm.tsx` - update NHTSA comment (line 510) - `frontend/src/pages/GuidePage/sections/VehiclesSection.tsx` - change "NHTSA database" to "vehicle database" - `frontend/src/pages/GuidePage/sections/SubscriptionSection.tsx` - change "NHTSA database" to "vehicle database" **Test files**: - `backend/src/features/vehicles/tests/integration/vehicles.integration.test.ts` - update vin_cache references if needed - Add unit tests for `VehiclesService.mapVinDecodeResponse()` mocking OcrClient - Add unit test for `OcrClient.decodeVin()` method **Regression verification**: - Confirm `POST /api/ocr/extract/vin` (image-based VIN extraction) still works - run existing OCR tests, verify PaddleOCR path untouched ### Milestone Dependencies ``` M1 (OCR endpoint) -> M2 (backend client) -> M3 (rewire controller + types + rename) -> M4 (delete NHTSA, must happen after M3 moves all code) -> M5 (frontend + tests, after M3 stabilizes API) ``` ### Files Summary | Action | Count | Details | |--------|-------|---------| | DELETE | 3 | `vehicles/external/nhtsa/` directory | | NEW | 3 | `ocr/app/routers/decode.py`, `ocr/tests/test_vin_decode.py`, `007_truncate_vin_cache.sql` | | MODIFY | ~20 | OCR engine/models/main/routers, OcrClient/types, vehicles controller/service/types/routes, platform models/README, 8 CLAUDE.md files, 7 frontend files, test files | *Verdict*: AWAITING_REVIEW | *Next*: QR plan-code, QR plan-docs
egullickson commented 2026-02-19 03:29:02 +00:00
Author
Owner
Copy Link

Plan v3: Revisions from QR plan-code and QR plan-docs

Phase: Plan-Review | Agent: Planner | Status: APPROVED

Addresses 4 RULE 0/1 findings from QR plan-code and 7 missing files from QR plan-docs.

Changes from v2

RULE 0 FIX: Remove TRUNCATE migration, use format-aware cache reads

Problem: TRUNCATE vin_cache is destructive with no rollback if Gemini is unavailable post-deploy.

Fix: Remove 007_truncate_vin_cache.sql. Instead, getVinCached() validates that raw_data has a success field (present in VinDecodeResponse, absent in old NHTSADecodeResponse). If format mismatch, treat as cache miss. Old NHTSA entries expire naturally via 1-year TTL. No destructive migration needed.

RULE 0 FIX: OCR decode router error status codes

Problem: No specified HTTP status for Gemini unavailability vs processing error.

Fix: OCR /decode/vin router returns:

  • 400 for invalid VIN format
  • 422 for GeminiProcessingError (Gemini returned bad data)
  • 503 for GeminiUnavailableError (credentials missing, SDK not installed)

Backend OcrClient.decodeVin() propagates these status codes. Controller maps: 503 -> 502 "VIN decode service unavailable", 422 -> 502 "VIN decode failed", 400 -> 400 "Invalid VIN".

RULE 0 FIX: Explicit VIN validation before cache/OCR

Problem: Gemini will hallucinate data for malformed VINs.

Fix: Controller decodeVin() step 1 explicitly validates VIN format using regex ^[A-HJ-NPR-Z0-9]{17}$ BEFORE cache read and before OCR call. Returns 400 immediately on failure. This matches the existing NHTSAClient.validateVin() behavior.

RULE 0 FIX: Cache write uses ON CONFLICT upsert

Fix: saveVinCache() uses INSERT INTO vin_cache ... ON CONFLICT (vin) DO UPDATE - same idempotent pattern as existing NHTSAClient.saveToCache().

RULE 1 FIX: Type transition safety

Problem: VinCacheEntry.rawData type change could break at milestone boundaries.

Fix: M3 is atomic - it extracts types from nhtsa.types.ts, changes rawData type, AND rewires the controller/service in a single milestone. The old NHTSAClient code continues to exist (deleted in M4) but is no longer imported by any file after M3 completes. No intermediate broken state.

RULE 2 NOTE: GenerationConfig per method

Fix: decode_vin() creates GenerationConfig locally within the method body (not cached as instance state). The existing self._generation_config for maintenance extraction is untouched.

RULE 2 NOTE: matchField() reuse confirmed

mapVinDecodeResponse() delegates to the same private matchField() method. No new fuzzy matching logic.

Additional files added to M3 (config/code)

These files have active vpic/NHTSA references that are functional, not just documentation:

  • backend/src/core/config/config-loader.ts - remove vpic from Zod schema in external config object (lines 46-48). RULE 0: failing to remove this causes config validation error at startup when vpic config is removed from YAML files
  • backend/src/core/config/feature-tiers.ts - update upgradePrompt from "NHTSA database" to "vehicle database" (line 32)
  • config/app/ci.yml - remove vpic: block (lines 26-28)
  • config/app/production.yml.example - remove vpic_api_url (line 25)
  • config/shared/production.yml - remove vpic: block (lines 110-111)

Additional files added to M4 (docs)

  • docs/PLATFORM-SERVICES.md - update "NHTSA vPIC API fallback with circuit breaker" to "Gemini VIN decode via OCR service" (line 38)
  • docs/USER-GUIDE.md - update "NHTSA database" references to "vehicle database" (lines 647, 658)
  • docs/TESTING.md - update vPIC mock references to OCR service mocks (lines 77, 197, 322-323)
  • backend/src/features/vehicles/external/README.md - remove NHTSAClient pattern reference (line 18)

Updated milestone file counts

Milestone Files (v2) Files (v3) Delta
M1 (#224) 6 6 --
M2 (#225) 2 2 --
M3 (#226) 5 10 +5 (config files)
M4 (#227) 12 16 +4 (docs files)
M5 (#228) ~12 ~12 --

Removed from plan

  • 007_truncate_vin_cache.sql - no longer needed (format-aware cache reads instead)

Verdict: APPROVED | Next: Create branch, begin execution

## Plan v3: Revisions from QR plan-code and QR plan-docs **Phase**: Plan-Review | **Agent**: Planner | **Status**: APPROVED *Addresses 4 RULE 0/1 findings from QR plan-code and 7 missing files from QR plan-docs.* --- ### Changes from v2 #### RULE 0 FIX: Remove TRUNCATE migration, use format-aware cache reads **Problem**: TRUNCATE vin_cache is destructive with no rollback if Gemini is unavailable post-deploy. **Fix**: Remove `007_truncate_vin_cache.sql`. Instead, `getVinCached()` validates that `raw_data` has a `success` field (present in `VinDecodeResponse`, absent in old `NHTSADecodeResponse`). If format mismatch, treat as cache miss. Old NHTSA entries expire naturally via 1-year TTL. No destructive migration needed. #### RULE 0 FIX: OCR decode router error status codes **Problem**: No specified HTTP status for Gemini unavailability vs processing error. **Fix**: OCR `/decode/vin` router returns: - `400` for invalid VIN format - `422` for `GeminiProcessingError` (Gemini returned bad data) - `503` for `GeminiUnavailableError` (credentials missing, SDK not installed) Backend `OcrClient.decodeVin()` propagates these status codes. Controller maps: 503 -> 502 "VIN decode service unavailable", 422 -> 502 "VIN decode failed", 400 -> 400 "Invalid VIN". #### RULE 0 FIX: Explicit VIN validation before cache/OCR **Problem**: Gemini will hallucinate data for malformed VINs. **Fix**: Controller `decodeVin()` step 1 explicitly validates VIN format using regex `^[A-HJ-NPR-Z0-9]{17}$` BEFORE cache read and before OCR call. Returns 400 immediately on failure. This matches the existing `NHTSAClient.validateVin()` behavior. #### RULE 0 FIX: Cache write uses ON CONFLICT upsert **Fix**: `saveVinCache()` uses `INSERT INTO vin_cache ... ON CONFLICT (vin) DO UPDATE` - same idempotent pattern as existing `NHTSAClient.saveToCache()`. #### RULE 1 FIX: Type transition safety **Problem**: `VinCacheEntry.rawData` type change could break at milestone boundaries. **Fix**: M3 is atomic - it extracts types from nhtsa.types.ts, changes `rawData` type, AND rewires the controller/service in a single milestone. The old NHTSAClient code continues to exist (deleted in M4) but is no longer imported by any file after M3 completes. No intermediate broken state. #### RULE 2 NOTE: GenerationConfig per method **Fix**: `decode_vin()` creates `GenerationConfig` locally within the method body (not cached as instance state). The existing `self._generation_config` for maintenance extraction is untouched. #### RULE 2 NOTE: matchField() reuse confirmed `mapVinDecodeResponse()` delegates to the same `private matchField()` method. No new fuzzy matching logic. --- ### Additional files added to M3 (config/code) These files have active vpic/NHTSA references that are functional, not just documentation: - `backend/src/core/config/config-loader.ts` - remove `vpic` from Zod schema in `external` config object (lines 46-48). **RULE 0**: failing to remove this causes config validation error at startup when vpic config is removed from YAML files - `backend/src/core/config/feature-tiers.ts` - update `upgradePrompt` from "NHTSA database" to "vehicle database" (line 32) - `config/app/ci.yml` - remove `vpic:` block (lines 26-28) - `config/app/production.yml.example` - remove `vpic_api_url` (line 25) - `config/shared/production.yml` - remove `vpic:` block (lines 110-111) ### Additional files added to M4 (docs) - `docs/PLATFORM-SERVICES.md` - update "NHTSA vPIC API fallback with circuit breaker" to "Gemini VIN decode via OCR service" (line 38) - `docs/USER-GUIDE.md` - update "NHTSA database" references to "vehicle database" (lines 647, 658) - `docs/TESTING.md` - update vPIC mock references to OCR service mocks (lines 77, 197, 322-323) - `backend/src/features/vehicles/external/README.md` - remove NHTSAClient pattern reference (line 18) ### Updated milestone file counts | Milestone | Files (v2) | Files (v3) | Delta | |-----------|-----------|-----------|-------| | M1 (#224) | 6 | 6 | -- | | M2 (#225) | 2 | 2 | -- | | M3 (#226) | 5 | 10 | +5 (config files) | | M4 (#227) | 12 | 16 | +4 (docs files) | | M5 (#228) | ~12 | ~12 | -- | ### Removed from plan - `007_truncate_vin_cache.sql` - no longer needed (format-aware cache reads instead) *Verdict*: APPROVED | *Next*: Create branch, begin execution
egullickson commented 2026-02-19 03:48:04 +00:00
Author
Owner
Copy Link

Milestone: M1-M3 Complete

Phase: Execution | Agent: Developer | Status: IN_PROGRESS

Completed

  • M1 (#224): OCR Python VIN decode endpoint - POST /decode/vin with GeminiEngine.decode_vin(), VinDecodeResponse model, VIN validation (400/422/503 error codes), 13 tests passing
  • M2 (#225): Backend OCR client - VinDecodeResponse type, OcrClient.decodeVin() method (JSON POST, not multipart)
  • M3 (#226): Rewire vehicles controller - Replaced NHTSAClient with OcrClient, moved cache logic to VehiclesService with format-aware reads (Gemini vs legacy NHTSA entries), renamed nhtsaValue->sourceValue, removed vpic config from Zod schema and YAML files. TypeScript compiles clean.

Remaining

  • M4 (#227): Remove NHTSA code and update docs (16 files)
  • M5 (#228): Frontend updates and tests (~12 files)

Commits

  • a75f7b5 feat: add VIN decode endpoint to OCR Python service (refs #224)
  • 3cd6125 feat: add backend OCR client method for VIN decode (refs #225)
  • 5cbf9c7 feat: rewire vehicles controller to OCR VIN decode (refs #226)

Verdict: IN_PROGRESS | Next: M4 and M5 (parallel)

## Milestone: M1-M3 Complete **Phase**: Execution | **Agent**: Developer | **Status**: IN_PROGRESS ### Completed - **M1** (#224): OCR Python VIN decode endpoint - `POST /decode/vin` with GeminiEngine.decode_vin(), VinDecodeResponse model, VIN validation (400/422/503 error codes), 13 tests passing - **M2** (#225): Backend OCR client - VinDecodeResponse type, OcrClient.decodeVin() method (JSON POST, not multipart) - **M3** (#226): Rewire vehicles controller - Replaced NHTSAClient with OcrClient, moved cache logic to VehiclesService with format-aware reads (Gemini vs legacy NHTSA entries), renamed nhtsaValue->sourceValue, removed vpic config from Zod schema and YAML files. TypeScript compiles clean. ### Remaining - **M4** (#227): Remove NHTSA code and update docs (16 files) - **M5** (#228): Frontend updates and tests (~12 files) ### Commits - `a75f7b5` feat: add VIN decode endpoint to OCR Python service (refs #224) - `3cd6125` feat: add backend OCR client method for VIN decode (refs #225) - `5cbf9c7` feat: rewire vehicles controller to OCR VIN decode (refs #226) *Verdict*: IN_PROGRESS | *Next*: M4 and M5 (parallel)
egullickson added
status
review
 and removed
status
in-progress
 labels 2026-02-19 03:53:25 +00:00
egullickson commented 2026-02-19 03:53:33 +00:00
Author
Owner
Copy Link

Milestone: All Milestones Complete - PR Open

Phase: Review | Agent: Developer | Status: AWAITING_REVIEW

All 5 Milestones Complete

  • M1 (#224): OCR Python VIN decode endpoint (7 files, 13 tests)
  • M2 (#225): Backend OCR client method (2 files)
  • M3 (#226): Rewire vehicles controller (9 files)
  • M4 (#227): Remove NHTSA code + docs (16 files, 3 deleted)
  • M5 (#228): Frontend + tests (7 files)

Quality Checks

  • Backend TypeScript: 0 errors
  • Frontend TypeScript: 0 errors
  • Backend lint: 0 errors
  • Frontend lint: 0 errors
  • OCR Python tests: 31 pass (13 new + 18 existing)
  • NHTSA grep: 0 references in code (only legacy cache format comments in service)

PR

PR #229 opened targeting main. Fixes #223, #224, #225, #226, #227, #228.

Verdict: AWAITING_REVIEW | Next: Quality Agent RULE 0/1/2 review

## Milestone: All Milestones Complete - PR Open **Phase**: Review | **Agent**: Developer | **Status**: AWAITING_REVIEW ### All 5 Milestones Complete - **M1** (#224): OCR Python VIN decode endpoint (7 files, 13 tests) - **M2** (#225): Backend OCR client method (2 files) - **M3** (#226): Rewire vehicles controller (9 files) - **M4** (#227): Remove NHTSA code + docs (16 files, 3 deleted) - **M5** (#228): Frontend + tests (7 files) ### Quality Checks - Backend TypeScript: 0 errors - Frontend TypeScript: 0 errors - Backend lint: 0 errors - Frontend lint: 0 errors - OCR Python tests: 31 pass (13 new + 18 existing) - NHTSA grep: 0 references in code (only legacy cache format comments in service) ### PR PR #229 opened targeting main. Fixes #223, #224, #225, #226, #227, #228. *Verdict*: AWAITING_REVIEW | *Next*: Quality Agent RULE 0/1/2 review
egullickson added
status
done
 and removed
status
review
 labels 2026-02-20 15:04:14 +00:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
status
backlog
status/blocked
status
done
status
in-progress
status
ready
status
review
type
bug
type
chore
type
docs
type
feature
No Label
status
done
type
feature
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
egullickson
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: egullickson/motovaultpro#223
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?