RBI GuideBot

An autonomous, agentic email-based assistant for NBFC professionals. It automatically ingests, classifies, and responds to queries about RBI guidelines and internal lending processes. It also includes a traditional web-based RAG chatbot interface.

Project Structure

• backend/: Contains the FastAPI server for the web chat and the complete agentic email workflow.
  • agents/: Houses the Classifier, Coordinator, and Retrieval agents.
  • core/: Core logic for orchestration, configuration, and the main agent service runner.
  • db/: Database models, session management, and CRUD operations for logging.
  • services/: Connectors for external services like Email (IMAP/SMTP).
  • api/: The FastAPI application for the web chat interface.
• frontend/: React.js UI for the web chat and feedback.
• data/: Source documents like RBI circulars, PDFs, etc.

Setup Instructions

1. Install Python 3.10+ and Node.js.
2. Create a .env file in the backend directory. Populate it with your credentials. See backend/core/config.py for all required variables (OpenAI, Pinecone, IMAP, SMTP).
3. Backend: cd backend && pip install -r requirements.txt
4. Frontend: cd frontend && npm install
5. Run Agentic Service: python -m backend.run_agent_service
6. Run Web Chat Backend: uvicorn backend.api.main:app --reload
7. Run Web Chat Frontend: cd frontend && npm start

Features

• Agentic Email Workflow:
  • Email Ingestion Agent: Automatically reads unread emails from a specified inbox (e.g., compliance@nbfcbot.com).
  • Email Classifier Agent: Uses an LLM to classify email intent (RBI Regulation, NBFC Lending, Combined, or Other).
  • Coordinator & Retrieval Agents: Routes the query to the correct RAG pipeline, retrieving context from separate vector databases (one for RBI rules, one for internal lending processes).
  • Response Generation Agent: Generates a contextual, natural-language answer with sources.
  • Email Reply Agent: Automatically sends the formatted response back to the customer.
• Database Logging: Every email interaction is logged in a database for auditing and review.
• Web Chat Interface: A traditional RAG chatbot interface is also available for direct queries.
• Document Ingestion: Scripts to ingest, chunk, embed, and store documents in Pinecone vector stores.

Technical Architecture

Architecture Add-on (Agentic Layer)

               ┌────────────┐
               │   Email    │
               │  Inbox     │
               └────┬───────┘
                    ▼
           ┌──────────────────┐
           │ Email Ingestion  │
           └────┬─────────────┘
                ▼
     ┌─────────────────────────────┐
     │ Email Classifier Agent      │
     │ (RBI / Lending / Other?)    │
     └────┬─────────────┬──────────┘
          ▼             ▼
   ┌────────────┐ ┌──────────────┐
   │ Regulation │ │   Lending    │
   │   Agent    │ │    Agent     │
   │ (RBI DB)   │ │ (NBFC DB)    │
   └────┬───────┘ └─────┬────────┘
        ▼               ▼
     ┌─────────────────────┐
     │ Coordinator Agent   │
     │ (Merge, Format)     │
     └────┬────────────────┘
          ▼
 ┌────────────────────┐
 │ Email Reply Agent  │
 └────────────────────┘

Web Chatbot Architecture

        ┌──────────────┐
        │   User       │
        │ (NBFC Staff) │
        └────┬─────────┘
             │
             ▼
 ┌─────────────────────┐
 │  Chatbot Frontend   │
 │ (React.js UI)       │
 └────┬────────────────┘
      ▼
 ┌────────────────────────────────────┐
 │ Backend API Server (FastAPI)       │
 └────┬────────────┬─────────────────┘
      ▼            ▼
┌────────────┐ ┌─────────────────────┐
│  RAG Flow  │ │  Chart Generator    │
│ (LLM + DB) │ │ (Matplotlib)        │
└────┬───────┘ └─────────────────────┘
     ▼
┌─────────────────────────────┐
│ Vector DB (Pinecone)        │
│  (Chunks + Embeddings)      │
└────────────┬────────────────┘
             ▼
   ┌──────────────────────┐
   │  Document Store      │
   │  (PDFs/Text)         │
   └──────────────────────┘

Sequence Diagram

Web Chatbot Flow

User → Frontend: Enter query
Frontend → Backend: Send query
Backend → Pinecone: Embed & search chunks
Backend → LLM: Generate answer
Backend → Chart Generator: (if numeric) Generate chart
Backend → SQL DB: Store search history/feedback
Backend → Frontend: Return answer, sources, chart
Frontend → User: Display results

Development Phases

1. Data Collection
2. Text Extraction, Chunking, Embedding
3. RAG Integration
4. Chatbot Interface + Backend API
5. Chart Generator Integration
6. Feedback Loop + Logging
7. Dockerization + Deployment