---

Tip

🏆 Featured in UptimeRobot's Global Case Study

"How Ambuj Kumar Tripathi built an 11-Node Agentic RAG system with 99.9% uptime on $0 infrastructure" This project was officially featured by UptimeRobot for its resilient production architecture on 512MB RAM, dual-purpose cold-start elimination, and hallucination-resistant document retrieval.

📖 Read the Full Published Spotlight on UptimeRobot Blog →

---

⚡ What Is This?

An autonomous, 11-node Agentic RAG pipeline that parses and queries complex Indian financial & legal documents — Union Budget, Finance Bill, Tax Laws, PF/Pension Schemes, RBI KYC, and Constitution of India — using a purpose-built state machine that thinks before it answers.

Unlike traditional RAG (retrieve → generate), this system employs an agentic flow where each query passes through specialized nodes that classify intent, cross-question vague queries, guard against hallucinations, and verify answer grounding — all orchestrated via LangGraph StateGraph.

---

🌿 Branches

Branch	Description
main	Production — stable, lean version live on Render (512MB RAM constraints)
v2-local-heavy	Parallel Vector Retrieval + Cohere Neural Reranking. 👉 View Architecture

---

🏗️ System Architecture

🔮 11-Node LangGraph StateGraph — Animated Architecture

_{✨ Classifier → 6-Path Routing → Retrieval → Rerank → Generate → Hallucination Guard → Post-Process}

---

🧠 The 11-Node Agentic RAG Pipeline

Node	Purpose	Key Detail
1. Classifier	Intent detection + 6-path routing	Returns structured JSON: intent · doc_type · confidence · search_intents
2. Reject	Safety guard	Blocks abusive + jailbreak queries with regex blocklist guardrail
3. Greet	Efficiency bypass	Handles greetings without hitting vector DB (zero cost)
4. CrossQuestioner	HITL clarification	Asks clarifying questions for vague queries (max 2 rounds)
5. Retriever	Dual vector search	Jina MRL → Pinecone → Parent-Child Resolution → Cohere Rerank Top 10
6. Web Search	Out-of-scope fallback	Tavily API — only fires after HITL user permission
7. Stock Tool	Native LLM tool calling	Gemini functionDeclarations + yfinance for live market data
8. Generator	LLM synthesis	Gemini 3.5 Flash Lite (primary) with pybreaker circuit breakers
9. HallucinationGuard	Answer verification	LLM-as-Judge — advisory mode (appends disclaimer, doesn't block)
10. PostProcess	Persistence + streaming	MongoDB + Redis cache + Langfuse tracing + SSE stream
11. Fallback	Circuit breaker recovery	pybreaker pattern: 3 API failures → graceful fallback message

---

🔧 Tech Stack

Category	Technology	Purpose
RAG Engine	LangGraph StateGraph	11-node autonomous state machine orchestration
	Jina v3 (MRL)	Matryoshka Representation Learning embeddings
	Cohere Neural Reranker	Advanced Stage-2 semantic filtering (V2)
	LlamaParse	LLM-native 3-tier document parsing
	Tavily Search API	Live Web Search fallback for Out-of-Scope queries
Backend & APIs	FastAPI + Uvicorn	Async REST API with SSE streaming
	Authlib + PyJWT	Google OAuth 2.0 + JWT session management
	WhatsApp Meta Cloud API	Real-time user bot interaction via Webhooks
Frontend	React 19 + Vite	SPA with lazy loading, dark theme, real-time streaming UI
Data Layer	Pinecone Serverless	14,662 vectors — core brain + ephemeral user uploads
	Supabase (PostgreSQL)	Parent chunk storage + file registry
	MongoDB (Motor)	Async chat history, feedback, user sessions
	Upstash Redis	Semantic caching (<100ms) + rate limiting + analytics
Reliability	Pybreaker	Circuit breaker pattern — 3 failures → auto-open → 30s reset
	Langfuse	Distributed tracing — LLM latency, token usage, cost tracking
	UptimeRobot	GET/HEAD health monitoring — zero cold starts
Deployment	Docker (Multi-stage) + Render	Frontend build → backend image → production serve

---

📊 Infrastructure Scale

Metric	Value
Total Chunks	15,408 (Financial Parser Portfolio)
Live Vectors	14,662 high-dimensional vectors in Pinecone (256d MRL)
Documents Indexed	20+ Indian Government Acts & Financial Frameworks
Parent Chunks	Stored in Supabase for full-context retrieval
Cache Latency	<100ms (Upstash Redis semantic cache)
Rate Limit	10 queries/min per user (Redis sliding window)
Session TTL	24h auto-cleanup (MongoDB TTL indexes)

---

📄 Documents Indexed

Category	Documents
Financial	Union Budget 2024-25, Finance Bill 2024-25, Income Tax Amendments
Pension/PF	EPF Scheme 1952, EPS Pension Scheme 1995, PMVVY, APY
Banking	RBI KYC Master Direction 2016, UPI Guidelines
Legal	Constitution of India, Consumer Protection Act

---

🔐 Security Architecture

7-Layer Upload Security Framework
─────────────────────────────────
Layer 1 │ Frontend Gating     │ .pdf only, 10MB limit, accept='.pdf'
Layer 2 │ Magic Byte Verify   │ %PDF- header validation (anti-spoofing)
Layer 3 │ Rate Limiting       │ 5 uploads/day per user+IP (Redis)
Layer 4 │ SHA-256 Dedup       │ Content-hash prevents re-indexing identical files
Layer 5 │ Session Isolation   │ is_temporary: true — auto-deletes on logout
Layer 6 │ TTL Auto-Cleanup    │ MongoDB 24h TTL on chunks + temp_uploads
Layer 7 │ Auth Guard          │ JWT verification on every API endpoint

---

🚀 Quick Start

Prerequisites

• Python 3.11+
• Node.js 18+
• API Keys: OpenRouter, Pinecone, MongoDB, Supabase, Google OAuth

Local Development

# Clone
git clone https://github.com/Ambuj123-lab/agentic-rag-financial-parser.git
cd agentic-rag-financial-parser

# Backend
python -m venv venv && venv\Scripts\activate  # Windows
pip install -r requirements.txt
cp .env.example .env  # Fill in your API keys
uvicorn app.main:app --reload

# Frontend (new terminal)
cd frontend
npm install && npm run dev

Docker (Production)

docker build -t financial-parser .
docker run -p 8000:8000 --env-file .env financial-parser

---

📁 Project Structure

agentic-rag-financial-parser/
├── app/
│   ├── main.py              # FastAPI app + SPA serving + health check
│   ├── api/
│   │   ├── auth.py           # Google OAuth + JWT + dev-login
│   │   ├── oauth.py          # Authlib Google client config
│   │   └── upload.py         # 7-layer secure file upload
│   ├── core/
│   │   └── config.py         # Pydantic Settings (env vars)
│   ├── db/
│   │   ├── mongodb.py        # Async Motor client + indexes
│   │   ├── pinecone_client.py # Pinecone Serverless init
│   │   └── supabase_client.py # Supabase PostgreSQL client
│   └── rag/
│       ├── graph.py          # ⭐ 11-Node LangGraph StateGraph
│       ├── routes.py         # Chat endpoints + SSE streaming
│       ├── embedder.py       # Jina v3 MRL embeddings
│       └── chunker.py        # Markdown + recursive splitting
├── frontend/
│   ├── src/
│   │   ├── pages/            # Landing, Dashboard, Admin, AuthCallback
│   │   ├── context/          # AuthContext (JWT state)
│   │   └── api/              # Axios client with interceptors
│   └── vite.config.js        # Dev proxy + code splitting
├── Dockerfile                # Multi-stage: Node build → Python serve
├── requirements.txt          # Pinned Python dependencies
└── .dockerignore             # Minimal Docker context

---

🌐 Live Links

Resource	URL
🚀 Live Application	agentic-rag-financial-parser.onrender.com
📖 RAG Documentation	ambuj-rag-docs.netlify.app
👤 Portfolio	ambuj-ai-portfolio.vercel.app
💻 Source Code	GitHub Repository

---

👨‍💻 Author

Ambuj Kumar Tripathi GenAI Engineer & RAG Systems Specialist | LLMOps

---

📜 License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).
Due to the integration of PyMuPDF (which is licensed under AGPL-3.0) for high-performance PDF parsing, this repository inherits the AGPL-3.0 license to comply with open-source copyleft requirements.

---

_{Built with 🧠 LangGraph • ⚡ FastAPI • ⚛️ React • 🔍 Pinecone • 🐘 Supabase • 🍃 MongoDB • 🔴 Redis}

7