Kontext Agent

Full-stack meeting and knowledge assistant with real-time transcription, periodic and final summaries, RAG search, and voice/chat.

What's Inside

• Backend: FastAPI (app/), Postgres for auth/chat/meeting data (transcripts and summaries encrypted at rest), ChromaDB (chroma_db/) for vectors.
• LLM: Groq chat completions with configurable model and temperature; PII is scrubbed before prompts.
• Speech: Deepgram STT (meetings and voice chat) and TTS.
• Frontend: React 19 + Vite + Tailwind (frontend/).
• Integrations: SMTP email, Notion export, JWT auth with signup/login/reset, rate limiting, meeting delete, and retention/TTL.

Core Features

• Live meeting WebSocket (/meeting/ws): PCM16 audio (binary or base64), interim/final transcripts, keep-alive plus auto reconnect, STOP triggers a structured final summary, runtime reconfig via {"type":"config","sample_rate":16000}.
• Summaries: Periodic delta-based snapshots (last N minutes) plus final Markdown summary; fallback summary if the LLM fails.
• Post-meeting chat (POST /meeting/{id|recent|any|latest}/chat): RAG over the meeting and cross-meeting context; accepts voice input; commands to email or Notion-export the summary.
• Document chat (POST /docs/chat): PDF/TXT upload (50 MB) with ingestion jobs, per-tab collections, optional RAG bypass, voice input.
• General chat (POST /chat/query): Remembers recent history and uses doc context when available; optional voice input.
• Voice chat socket (/voice-chat/ws): Real-time STT -> LLM -> TTS with conversation recall.
• Data hygiene: PII redaction before LLM calls, encrypted DB columns for transcripts/summaries, meeting retention/TTL, and hard delete.

Environment

Copy .env.example to .env and set at least:

• DATABASE_URL (async SQLAlchemy URI; Postgres via asyncpg recommended)
• GROQ_API_KEY
• DEEPGRAM_API_KEY
• JWT_SECRET (>=32 chars in production)
• ENCRYPTION_KEY (falls back to JWT_SECRET, but set a dedicated key for production)
• PRELOAD_EMBEDDINGS (set to true to download the embedding model at startup; defaults to false to speed container health checks)

Common optional keys:

• GROQ_MODEL, LLM_TEMPERATURE, LLM_TIMEOUT
• EMBEDDING_MODEL (defaults to all-MiniLM-L6-v2)
• CHROMA_DIR
• MEETING_RETENTION_DAYS (default 90), PERIODIC_SUMMARY_LOOKBACK_MINUTES (default 10)
• SMTP_HOST/PORT/USER/PASS for email
• NOTION_TOKEN, NOTION_PAGE_ID
• CORS_ORIGINS, FRONTEND_URL, RATE_LIMIT_REQUESTS, RATE_LIMIT_PERIOD

Local Development

Backend (Python 3.11):

python -m venv .venv
.\.venv\Scripts\activate      # PowerShell
pip install -r requirements.txt
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Frontend (Node 20):

cd frontend
npm install
npm run dev

• API base: http://localhost:8000
• Frontend dev: http://localhost:5173

Docker

Quick start (prebuilt images)

Prebuilt images are on Docker Hub:

• Backend: lazyghost1/kontext-backend
• Frontend: lazyghost1/kontext-frontend

Place a .env (copy from .env.example) in the repo root so the backend has your keys/secrets.

Save this as docker-compose.yml (or adapt your existing one):

services:
  postgres:
    image: postgres:16-alpine
    environment:
      POSTGRES_USER: postgres
      POSTGRES_PASSWORD: postgres
      POSTGRES_DB: kontext_agent
    volumes:
      - postgres_data:/var/lib/postgresql/data

  backend:
    image: lazyghost1/kontext-backend:latest
    env_file: .env
    environment:
      APP_ENV: production
      APP_HOST: 0.0.0.0
      APP_PORT: "8000"
      DATABASE_URL: postgresql+asyncpg://postgres:postgres@postgres:5432/kontext_agent
      CHROMA_DIR: /app/chroma_db
      PRELOAD_EMBEDDINGS: "true"
      CORS_ORIGINS: http://localhost
    ports:
      - "8000:8000"
    depends_on:
      - postgres
    volumes:
      - chroma_data:/app/chroma_db

  frontend:
    image: lazyghost1/kontext-frontend:latest
    environment:
      VITE_API_BASE_URL: http://localhost:8000
      VITE_WS_URL: ws://localhost:8000
    ports:
      - "80:80"
    depends_on:
      - backend

volumes:
  postgres_data:
  chroma_data:

Run it:

docker compose up -d

Build locally (optional)

If you prefer to build from source:

docker compose up --build

API Quick Reference

• GET / and GET /health
• POST /auth/signup | /login | /forgot-password | /reset-password
• WS /meeting/ws - send audio; STOP to finalize; ACTION: EMAIL <addr> or ACTION: NOTION to export
• GET /meeting/history, /meeting/{id}/transcript, /meeting/{id}/summary
• DELETE /meeting/{id} (hard delete transcript, summaries, embeddings)
• POST /meeting/{id|recent|any|latest}/chat
• POST /chat/query, DELETE /chat/history
• POST /docs/upload, GET /docs/status/{job_id}, POST /docs/chat, DELETE /docs/clear
• WS /voice-chat/ws

Data and Storage

• Transcripts and summaries: encrypted in Postgres with retention (MEETING_RETENTION_DAYS).
• Vector store: chroma_db/ (Chroma persistent client).

Testing

pytest

Tests stub Chroma/RAG for speed; they use SQLite by default unless DATABASE_URL is set.

Contributing

Pull requests are welcome. Quick checklist:

• Fork the repo and create a branch (git checkout -b feat/your-idea).
• Backend: Python 3.11, python -m venv .venv && .\.venv\Scripts\activate, pip install -r requirements.txt, run pytest.
• Frontend: Node 20, cd frontend && npm install && npm run lint.
• Run apps locally (uvicorn app.main:app --reload --port 8000 and npm run dev). If you change ports, set VITE_API_BASE_URL (and VITE_WS_URL for sockets) in frontend/.env.local.
• Keep PRs focused; include screenshots for UI changes when helpful.
• In the PR description, note the motivation, what changed, and how you tested.

Troubleshooting

• Port already in use: stop the conflicting process (netstat -ano | findstr :8000) or change the port and update VITE_API_BASE_URL.
• Slow first start: embedding model download can take time; set PRELOAD_EMBEDDINGS=false to skip upfront (it will lazy-load later).

Security Notes

• Set strong JWT_SECRET and explicit CORS_ORIGINS in production.
• Do not commit real .env values.