Skip to content

activeContext.md

Latest commit

 

History

History
248 lines (206 loc) · 8.3 KB

activeContext.md

File metadata and controls

248 lines (206 loc) · 8.3 KB

Active Context - Demand Letter Generator

Current Focus

🚀 Project Initialization Complete! Ready for Milestone 1

We have successfully established the foundation for the Demand Letter Generator:

  1. ✅ Memory Bank established (all core documents)
  2. ✅ Project structure created (frontend, backend, ai-service)
  3. ✅ Initial codebase scaffolding complete
  4. ✅ Configuration files set up (Docker, TypeScript, Prisma)
  5. ✅ Git repository initialized

Next: Begin Milestone 1 - Infrastructure & Storage implementation

Recent Decisions

Tech Stack Confirmed (Nov 11, 2025)

  • Frontend: React + TypeScript (not Next.js)
  • Backend: Node.js with Express
  • AI Service: Python (AWS Lambda) + Anthropic API
  • Database: PostgreSQL
  • Real-time: TipTap + Y.js + WebSockets + Redis
  • Storage: AWS S3
  • Architecture: Microservices

PRD Evolution

  • Started with initial PRD (basic requirements)
  • Updated to PRD v2 with critical additions:
    • Fact approval workflow (human-in-the-loop quality gate)
    • Multi-document handling (up to 50 PDFs)
    • Single-firm deployment model (no multi-tenancy)
    • Detailed versioning and audit trail
    • Clear performance limits

Permission Model

  • Decision: All users can create/edit firm templates
  • Rationale: No admin role complexity in V1, trust-based single-firm model

Next Steps

Immediate (Now)

  1. ✅ Memory Bank files created
  2. ✅ Create project directory structure
  3. ✅ Initialize Git repository
  4. ✅ Set up frontend (React + TypeScript + Vite)
  5. ✅ Set up backend (Node.js + Express + TypeScript)
  6. ✅ Set up AI service (Python + Lambda structure)
  7. ✅ Create shared types directory
  8. ✅ Docker Compose for local Postgres + Redis
  9. ✅ Initial README with setup instructions
  10. ✅ Prisma schema defined
  11. ✅ Basic Express server with health check
  12. ✅ React app with routing structure

Milestone 1: Infrastructure & Storage

Goal: Database, S3, authentication foundation

Tasks:

  • Database schema implementation (Prisma or TypeORM)
  • Run initial migrations (users, documents, templates tables)
  • S3 bucket setup with encryption
  • JWT authentication implementation
    • Register endpoint
    • Login endpoint
    • Refresh token mechanism
    • Auth middleware
  • Basic health check endpoints
  • Error handling middleware
  • Request logging

Acceptance Criteria:

  • User can register and login
  • JWT tokens are issued and validated
  • Database tables created and accessible
  • S3 bucket ready for file uploads
  • Health check returns 200 OK

Milestone 2: PDF Intake Layer

Goal: Upload PDFs, extract text, store metadata

Tasks:

  • PDF upload endpoint (multipart/form-data)
  • S3 upload integration with encryption
  • Python service: text extraction from PDFs
  • Store extracted text in database
  • Frontend: PDF upload UI component
  • Frontend: Document list view
  • Error handling for invalid/corrupted PDFs

Acceptance Criteria:

  • User can upload multiple PDFs (up to 50)
  • Files stored in S3 with proper naming
  • Text extracted and stored in pdfs table
  • UI shows upload progress and success
  • Invalid files rejected with clear error messages

Milestone 3: Template System

Goal: Create templates with variables and paragraph modules

Tasks:

  • Template CRUD API endpoints
  • Paragraph module CRUD API endpoints
  • Template builder UI (variable editor)
  • Paragraph library UI (create, tag, search)
  • Drag-and-drop paragraph insertion
  • Template preview with sample data
  • Template validation (required fields, structure)

Acceptance Criteria:

  • User can create template with variables
  • User can create paragraph modules with tags
  • User can preview template with sample data
  • Templates saved and retrievable
  • Paragraph modules associated with templates

Milestone 4: AI Draft Generation Pipeline

Goal: Extract facts, approve, generate draft

Tasks:

  • Python Lambda: fact extraction from PDF text
  • Anthropic API integration (Claude 3.5 Sonnet)
  • Fact extraction prompt engineering
  • Fact approval UI (approve/edit/reject)
  • Multi-stage draft generation pipeline
    • Facts → Narrative
    • Narrative + Template → Structure
    • Structure + Tone → Draft
  • Draft generation API endpoint
  • Display generated draft in editor
  • Error handling for AI failures

Acceptance Criteria:

  • Facts extracted with PDF citations
  • User can approve/edit/reject each fact
  • Only approved facts used in generation
  • Draft generated in < 10 seconds
  • Draft displayed in TipTap editor
  • Graceful degradation if AI fails

Milestone 5: Collaborative Editing

Goal: Real-time multi-user editing

Tasks:

  • TipTap editor integration
  • Y.js CRDT setup
  • WebSocket server implementation
  • Redis pub/sub for multi-instance scaling
  • Presence indicators (active users)
  • Autosave (every 30 seconds)
  • Version history (snapshot on save)
  • Restore previous version functionality
  • Conflict-free editing (Y.js handles this)

Acceptance Criteria:

  • 2+ users can edit simultaneously
  • Changes sync in real-time (< 100ms latency)
  • Presence indicators show active users
  • Document autosaves without user action
  • User can view and restore previous versions
  • No data loss during network interruption

Milestone 6: Export

Goal: Export to DOCX with firm letterhead

Tasks:

  • DOCX generation library setup (docxtemplater or python-docx)
  • Firm letterhead template system
  • Export API endpoint
  • Convert TipTap JSON to DOCX format
  • Preserve formatting (bold, italic, lists, etc.)
  • Download endpoint for generated DOCX
  • Frontend: Export button + download handling
  • Audit log: record export events

Acceptance Criteria:

  • User clicks "Export" and receives DOCX file
  • File opens correctly in Microsoft Word
  • Formatting preserved (fonts, styles, spacing)
  • Firm letterhead applied
  • Export logged in audit trail

Active Questions & Considerations

Open Questions

  1. Firm Letterhead: Should letterhead be:

    • Uploaded as DOCX template file?
    • Configured via UI (logo + header/footer text)?
    • Hardcoded initially with config file?

  2. Session Management: How long should user sessions last?

    • Current plan: 15-min access token, 7-day refresh token
    • Should we have "remember me" option?

  3. PDF Preview: Should we render PDF thumbnails in UI?

    • Requires additional processing
    • Nice UX but not critical for V1

  4. Fact Extraction Quality: What's acceptable accuracy?

    • Need golden dataset for testing
    • May require prompt iteration

  5. Cost Monitoring: Should we track AI API costs per document?

    • Useful for analytics but adds complexity
    • Could add in post-V1

Technical Risks

  • Y.js Complexity: Real-time collaboration is technically challenging
    • Mitigation: Start with basic TipTap, add Y.js incrementally
  • AI Hallucinations: Facts may be inaccurate
    • Mitigation: Human approval workflow (already designed)
  • PDF Parsing: Not all PDFs extract cleanly
    • Mitigation: Clear error messages, manual text input fallback
  • WebSocket Scaling: Redis pub/sub required for multi-instance
    • Mitigation: Start with single instance, add Redis when needed

Design Decisions Needed

  1. Tone Slider (P1 Feature):

    • When user adjusts tone, regenerate entire draft or inline edits?
    • Decision: Defer to post-V1, focus on single tone generation first

  2. Track Changes (P1 Feature):

    • Redline mode like Word's track changes
    • Decision: Defer to post-V1, version history covers basic need

  3. OCR (P1 Feature):

    • AWS Textract for scanned PDFs
    • Decision: Defer to post-V1, focus on text-based PDFs

Current Blockers

None - ready to proceed with implementation.

Recent Changes

  • Memory Bank structure created
  • Project scope clarified through PRD updates
  • Tech stack confirmed with user

Collaboration Notes

  • Working with: User (Dohoon Kim)
  • Communication style: Direct, focused on essentials
  • Decision speed: Fast - user prefers action over extensive planning
  • Preferred approach: Build incrementally, iterate based on working software

Environment Notes

  • Workspace: /Users/dohoonkim/GauntletAI/Demand Letter/
  • Date: November 11, 2025
  • OS: macOS (darwin 24.6.0)
  • Shell: zsh