navidocs/CLOUD_SESSION_PROMPT_2_DOCUMENT_SEARCH.md

Cloud Session 2: Advanced Document Search

Session ID: CLOUD-2-DOCUMENT-SEARCH Timeline: 90 minutes Deadline: 4 hours from now (Riviera Plaisance presentation) Target: Ship OCR search + document classifier improvements

---

Your Mission

Improve the document search experience so boat owners can find maintenance manuals, insurance papers, warranty docs instantly. Current system works but needs:

• Sticky engagement: "Where's my engine manual?" → answers in 2 seconds
• OCR accuracy: Text extraction from boat documents (technical manuals, warranty cards, insurance papers)
• Smart grouping: Show warranty + insurance + service history together
• Auto-tagging: Classify documents by type (engine, electrical, safety equipment, etc.)

This prevents "Where's the engine manual?" crisis during mechanical emergencies.

---

Quick Start

1. Clone repo:

   git clone https://github.com/dannystocker/navidocs.git && cd navidocs
   

2. Read context:

   • OCR_PIPELINE_SETUP.md - Current OCR implementation (Tesseract + Google Vision)
   • BUILD_COMPLETE.md - What search features already work

3. Check Meilisearch status:

   curl http://localhost:7700/health
   

4. Review OCR API:

   grep -r "ocr\|vision" src/api/
   

---

Your Task List

• Diagnostic: Review current OCR results quality

  • Upload 5 test documents (warranty card, engine manual, insurance doc)
  • Check extraction quality (confidence scores, missing text?)
  • Document findings in SEARCH_QUALITY_REPORT.md

• Implement: Smart document classifier

  • Add documentType field to Document table (engine, electrical, hull, interior, warranty, insurance, service, safety)
  • Create classifier endpoint: POST /api/documents/classify (reads OCR text → returns type)
  • Support manual override (user selects type if AI wrong)

• Improve: Search results ranking

  • Boost warranty + service docs to top
  • Show document type icon + confidence score
  • Group results by document type

• Test: Search UX with real documents

  • "engine manual" → Find service manuals
  • "warranty" → Find all warranty cards + service plans
  • "electrical" → Find electrical system diagrams + parts docs

• API endpoints:

  • POST /api/documents/classify - Auto-classify document type
  • GET /api/documents/by-type/:type - Filter by type
  • GET /api/search/advanced - Enhanced search with type + relevance ranking

• Git commit: [AGENT-2] Add document classifier and search ranking

• Create issue: [AGENT-2] DEPLOY-READY: Document Search Improvements with:

  • Test results (5 documents, accuracy %)
  • Search quality report
  • Performance metrics (search latency)
  • Deployment checklist

---

Technical Context

Current Stack:

• OCR Pipeline: Tesseract (local) + Google Vision API (backup)
• Search Engine: Meilisearch (localhost:7700)
• Database: PostgreSQL - Document table with content field (OCR extracted text)
• Frontend: Next.js search UI component

Key Files:

• src/api/ocr/route.ts - Current OCR implementation
• src/api/search/route.ts - Search endpoint
• src/components/DocumentSearch.tsx - Search UI
• prisma/schema.prisma - Document model

Design Specs:

• Document types: engine, electrical, hull, interior, warranty, insurance, service, safety, other
• OCR text stored in Document.content (PostgreSQL)
• Meilisearch index includes: title, type, confidence, upload_date
• Search results show: document title, type badge, 2-line preview, relevance score

---

Sample Test Documents

Create these for testing OCR quality:

1. Engine Manual - Technical specifications, maintenance schedule
2. Warranty Card - Registration, coverage terms, contact info
3. Insurance Document - Policy details, coverage limits
4. Service Record - Date, service performed, parts replaced
5. Electrical Diagram - System schematic with part numbers

Quality Thresholds:

• OCR confidence >85% = no review needed
• 70-85% = flag for manual review
• <70% = skip from search (mark as low-confidence)

---

Critical Notes

1. Boat owner pain point: Mechanical emergency at 2am, need engine manual NOW
2. Search must be fast: <500ms response time (cached results)
3. OCR accuracy matters: Wrong document type = wrong answers
4. Offline support: Downloaded documents searchable without internet
5. Mobile first: Search on small screens must work perfectly

---

GitHub Access

• Repo: https://github.com/dannystocker/navidocs
• Branch: feature/document-search (create from main)
• Base for PR: main branch

---

Success Criteria

✅ Document classifier working (type detection >80% accurate) ✅ Search results ranked by type + relevance ✅ Test documents fully searchable ✅ OCR quality report completed ✅ API endpoints tested and working ✅ No console errors ✅ Git commit with [AGENT-2] tag

---

If Blocked

1. Check Google Vision API credentials: echo $GOOGLE_VISION_API_KEY
2. Verify Tesseract installed: tesseract --version
3. Review current OCR: cat OCR_PIPELINE_SETUP.md
4. Check Meilisearch index: curl http://localhost:7700/indexes/documents/stats
5. Create blocker issue: [AGENT-2] BLOCKER: [description]

---

Reference Files

• OCR_PIPELINE_SETUP.md - Complete OCR setup guide
• ARCHITECTURE-SUMMARY.md - System architecture
• SMOKE_TEST_CHECKLIST.md - Testing procedures