How openchatai/openchat works
Project type and purpose
• This is a full-stack, multi-service chatbot console (“OpenChat”) that orchestrates several backends and frontends to let users spin up custom LLM-powered chatbots.
• The README states that the original Laravel backend (“backend-server”) has been largely rewritten in Python/Django (“dj_backend_server”), with an auxiliary Next.js/TypeScript “llm-server” for embedding and chat operations, plus a React-based widget under widgets/search.
File structure
• Top-level services:
• backend-server – legacy Laravel PHP REST/Web controllers & Blade views
• dj_backend_server – new Django REST + WebSocket (Channels) app, Celery tasks, admin, templates, static assets
• llm-server – Next.js API routes for chat/ingest, vector-DB logic in TS
• widgets/search – standalone React widget packaged via Vite
• Deployment configs: docker-compose.yml, service Dockerfiles, Makefiles
• Data/config: environment files (common.env, example.env), vector-DB config under llm-server/config
• Infra: nginx templates, Celery broker/worker, Django Channels routing
Key insights from the README
• Uses OpenAI (or Azure) for LLM calls
• Supports Pinecone or Qdrant vector databases
• Python Django is primary backend, with Celery for background ingestion
• Frontend will be migrated to Next.js/TS; for now, Django templates & the React widget co-exist
• Client UIs
– Browser clients hitting Django templates
– Embedded React widget (widgets/search) on customer sites
• Django Backend (dj_backend_server)
– Django REST API endpoints
– WebSocket consumers (Channels)
– Celery Workers for ingestion tasks
• LLM Microservice (llm-server)
– Next.js API routes (/api/ingest, /api/chat)
– Embedding & retrieval logic
• Data Stores
– Relational DB (Postgres/MySQL for Django models)
– Message broker (Redis/RabbitMQ for Celery)
– Vector store (Pinecone or Qdrant)
• External LLM APIs
– OpenAI (or Azure OpenAI)
1. User → Django REST or WebSocket
2. Django API → Celery (enqueue ingest/chat tasks)
3. Celery Worker → llm-server ingest endpoint (send documents or codebase)
4. llm-server → Vector Store (upsert embeddings)
5. User → Django API chat endpoint → llm-server chat endpoint →
Vector Store + OpenAI → llm-server response → Django → Client
6. Django ↔ Relational DB for user/chatbot metadata & chat histories
c. Architectural patterns & principles
• Microservice separation of concerns: Django for orchestration, Next.js for vector/LLM logic
• Event-driven ingestion via Celery tasks and listeners
• CQRS-like read/write separation (writes via ingestion pipeline, reads via vector retrieval)
• WebSocket pub/sub for real-time chat (Django Channels)
d. Key technologies & frameworks
• Python/Django, Django REST Framework, Django Channels, Celery
• Node.js/Next.js, TypeScript
• Pinecone or Qdrant vector DB
• Redis or RabbitMQ
• OpenAI/Azure OpenAI APIs
• Docker, Docker Compose, nginx