InsightRAG — Adaptive, Self-Corrective Retrieval-Augmented Generation System

A modular, production-ready RAG pipeline built with FastAPI, LangGraph, and Streamlit, featuring adaptive routing, self-grading, and hallucination correction for fact-grounded responses.

---

🚀 Overview

InsightRAG combines internal vector retrieval and external web search to deliver factual, explainable AI responses.
It integrates LangGraph for adaptive orchestration, LangSmith for tracing and metrics, and FAISS for fast semantic search.

The system supports three RAG paradigms:

• Adaptive RAG → Chooses between internal (FAISS) or external (web) retrieval.
• Self RAG → Grades document relevance and factual alignment.
• Corrective RAG → Regenerates hallucinated answers for better accuracy.

---

🧰 Tech Stack

Layer	Technologies
Backend	Python, FastAPI, LangGraph, LangChain
Models	Gemini, GPT-4, Claude
Vector Store	FAISS, HuggingFace Embeddings (all-MiniLM-L6-v2)
Web Search	Tavily API
Frontend	Streamlit
Observability	LangSmith Tracing & Metrics
Documentation	FastAPI Swagger/OpenAPI

For a detailed architectural diagram and agent interaction flow, check out the LangGraph Flow Diagram

📡 API Endpoints (FastAPI)

Endpoint	Method	Description	Request Schema
/api/v1/load_vectors	POST	Upload and embed text/JSON documents	UploadFile
/api/v1/query_vectors	POST	Query FAISS for similar docs	QueryRequest: {query: str, k?: int}
/api/v1/model_query	POST	Direct LLM query to model	ModelQueryRequest: {query: str, provider?: str, model?: str}
/api/v1/rag_query	POST	Run full Adaptive RAG pipeline	RAGQueryRequest: {query: str, k?: int}
/health	GET	Health check endpoint	-
/	GET	API root with version info	-

💻 Frontend (Streamlit)

A minimal chatbot UI to demo all endpoints. Launch it using:

streamlit run streamlit_app.py

🎯 Features

Core Capabilities

• Adaptive routing between vector store and web search
• Document relevance grading and factual alignment
• Hallucination detection and correction
• Multi-model support (Gemini, GPT-4, Claude)
• Efficient FAISS-based vector similarity search

Interface

• Interactive Streamlit chatbot interface
• Rich metadata and source attribution
• Support for both direct LLM and RAG queries
• Real-time tracing and monitoring via LangSmith
• Swagger/OpenAPI documentation

Data Handling

• Text and JSON document ingestion
• HuggingFace embeddings integration
• Persistent FAISS index storage
• Web search capability via Tavily
• Structured response formatting

🧪 Quick Start

1️⃣ Clone & Setup

git clone https://github.com/<yourname>/InsightRAG.git

cd InsightRAG

python -m venv venv

source venv/bin/activate   # or venv\Scripts\activate

pip install -r requirements.txt

2️⃣ Environment Variables (.env)

GOOGLE_API_KEY="<your_google_api_key>"
OPENAI_API_KEY="<your_openai_api_key>"
TAVILY_API_KEY="<your_tavily_api_key>"
LANGCHAIN_TRACING_V2="true"
LANGCHAIN_PROJECT="AdaptiveRAG"
LANGCHAIN_API_KEY="<your_langsmith_api_key>"

3️⃣ Run Services

# Start backend
uvicorn app.main:app --reload

# Start frontend
streamlit run frontend\streamlit_app.py

🧠 Example Query

curl -X POST "http://127.0.0.1:8000/api/v1/rag_query" \
-H "Content-Type: application/json" \
-d '{"query": "Explain retrieval augmented generation", "k": 3}'

📈 Metrics & Tracing

LangSmith Integration

• Comprehensive tracing of all agent nodes (Router, Retriever, Grader)
• Real-time monitoring and debugging via LangSmith dashboard
• Trace metadata includes: component, agent, model, index, k, query_len, num_docs, avg_doc_score

Observability

• Environment Setup: LANGCHAIN_TRACING_V2=true, LANGCHAIN_PROJECT=AdaptiveRAG
• LangSmith initialization at FastAPI startup
• Privacy-aware tracing: No PII or full document content in traces
• Performance metrics and latency tracking
• Error monitoring and alerting capabilities

🧭 Future Enhancements

• PDF and Docx Document Support: Add capability to process PDF documents and extract text for ingestion
• Conversation Memory: Implement chat history and context management
• Adding Data Visualisations Capablities