Sohel2309/enterprise-rag

Enterprise RAG Platform

Production-grade retrieval-augmented generation system with hybrid retrieval, intelligent reranking, and conversation memory.



Overview

Enterprise RAG Platform combines BM25 sparse + BGE dense embeddings with Cohere reranking to deliver accurate, context-aware answers. Features conversation memory, real-time document upload, and automated RAGAS evaluation.

📊 Key Results

🏗️ Architecture

Retrieval Pipeline:

• —BM25 (sparse) + BGE-large embeddings (dense) with RRF fusion
• —Cohere cross-encoder reranking (top-60 → top-10)
• —DeBERTa-v3 NLI hallucination detection per-chunk

Generation:

• —Groq openai/gpt-oss-120b with 3-turn conversation memory
• —Automatic session management + real-time document ingestion

🚀 Tech Stack

FastAPI · Streamlit · FAISS · BGE Embeddings · Cohere Reranking · Groq LLM · RAGAS Evaluation · HuggingFace Spaces

⚡ Quick Start

# Install
pip install -r requirements.txt

# Run locally
uvicorn api.rag_api:app --reload --port 8000  # Terminal 1

streamlit run ui/streamlit_app.py             # Terminal 2

# Visit http://localhost:8501

📁 Project Structure

Enterprise-RAG/
├── api/                 # FastAPI backend
├── ingestion/           # PDF/DOCX parsing + 5 chunking strategies
├── retrieval/           # BM25 + FAISS + Cohere reranking
├── generation/          # Groq LLM client
├── evaluation/          # RAGAS + hallucination detection
├── ui/                  # Streamlit dashboard (3 tabs)
└── data/
    └── processed/       # Pre-indexed chunks (FAISS + BM25)

🎯 Features

✅ Hybrid retrieval (BM25 + dense embeddings)

✅ Intelligent reranking with Cohere

✅ Multi-turn conversation memory

✅ Real-time PDF/DOCX upload

✅ Automated hallucination detection

✅ Interactive evaluation dashboard

📈 Benchmark

Fixed-size chunking outperformed sentence-based by 17% on context precision. Evaluated on 50 ArXiv papers with 355 QA pairs.

🔍 Evaluation Methodology & Limitations

The metrics in Key Results are produced by an LLM-as-judge evaluation, not a human-labeled benchmark. The judge model is openai/gpt-oss-120b, run against a held-out set of QA pairs using the exact same retrieval configuration as production (hybrid retrieval + Cohere reranking). The current reported run evaluated 45 valid samples.

Two limitations worth stating plainly:

• —Self-judging bias: the same model (openai/gpt-oss-120b) is used for both answer generation and evaluation scoring. LLM judges are known to show some self-preference bias, so these scores should be read as internal, project-level evidence of pipeline quality — not as an external or industry benchmark.
• —Sample size: 45 samples is directional evidence appropriate for a project of this scope, not a statistically rigorous evaluation.

These numbers are useful for demonstrating that retrieval and generation quality were measured and iterated on, not just assumed to work.

🔗 Links

• —Live Demo: https://sohel2309-enterprise-rag.hf.space
• —Code: https://github.com/Sohel2309/Enterprise-RAG

MIT License | Last Updated: June 2026