DocMind

Live Deployment: https://docmind1.netlify.app/

DocMind is a real-time, agentic Retrieval-Augmented Generation (RAG) platform. Rather than relying on a standard single-pass retrieval pipeline, DocMind executes a bounded autonomous loop that dynamically selects and executes tools based on user queries, streaming its execution state to the client in real time.

---

Architecture & Core Systems

• Bounded Agentic Loop: The core engine is a custom autonomous loop (max 5 iterations) powered by gpt-oss-20b via Groq.
• Real-Time Telemetry: Tool executions and intermediate results are streamed to the React client via Server-Sent Events (SSE) using fetch and ReadableStream. No black-box buffering.
• Strict Session Isolation: Fully stateless architecture. A session_id is generated client-side, injected via the X-Session-Id header, and securely scoped into every Supabase DB query. Identifiers never bleed into streamed events.
• Concurrency & Rate Limit Handling: Implements a Postgres trigger (enforce_session_document_cap) to close race conditions on the 3-document upload limit, paired with idempotent ingestion (hash-based) and exponential backoff wrappers (2s/4s/8s) for LLM rate limits.

---

Tech Stack

Layer	Technology
Frontend	React 18 + Vite (Deployed on Netlify)
Backend	FastAPI / Python 3.11 (Deployed on Render)
LLM Inference	openai/gpt-oss-20b via Groq (litellm)
Embeddings	Google gemini-embedding-2 (768-dim)
Vector DB	Supabase pgvector (Cosine Similarity Search)

---

Tooling & Capabilities

The agent has autonomous access to the following callable tools:

1. retrieve_documents: Semantic vector search over ingested chunks.
2. list_available_documents: Contextual awareness of the current session's state.
3. summarize_document: Full-text document synthesis.
4. compare_documents: Cross-document analysis and routing.

---

Local Environment Setup

Prerequisites

• Python 3.11+ & Node.js 18+
• Supabase instance (with pgvector enabled)
• Groq API Key & Google API Key (for embeddings)

Backend Pipeline

cd backend
cp .env.example .env 
# Configure SUPABASE_URL, SUPABASE_SERVICE_KEY, GROQ_API_KEY, GOOGLE_API_KEY

pip install -r requirements.txt
uvicorn main:app --reload

Frontend Client

cd frontend
cp .env.example .env 
# Set VITE_API_URL=http://localhost:8000

npm install
npm run dev

---

API Reference

Method	Endpoint	Function
GET	/health	System health check
POST	/upload	Ingest PDF/TXT (Requires X-Session-Id)
GET	/documents	List session-scoped documents
DELETE	/documents/{id}	Purge document from session
POST	/chat	Standard synchronous agent chat
POST	/chat/stream	SSE streaming endpoint

Note: The Render deployment utilizes the X-Accel-Buffering: no header on /chat/stream to prevent proxy buffering of SSE chunks.

---