PaperRAG

Click to view Demo video on youtube

---

A Production grade full stack RAG application for answering queries related to online research papers

Architecture

---

Query Router

• Routes the query on the basis of requirement of context

Query Enhancement Techniques

• Use Step-back prompting to capture broad range of topics for enriched context
• Use Context aware query decomposition to generate a quality queries using retreived context from step-back prompt

Advanced Retreival

• Filter retreived documents to remove redundant and non-relevant documents
• Rerank the retreived documents using cross-encoder to get more relevance of ordering of documents

Generation

• Generate the answer with given context only (if query router had pointed to rag portion) else generate only from learned patterns

Evaluation Summary

Comparison of RAG Approaches (Mean Metrics)

Metric	Naive RAG	Advanced RAG (WO Hybrid Retrieval)	Advanced RAG + Hybrid Retrieval
Correctness	0.188	0.215	0.285
Groundness	0.876	0.813	0.830
Relevance	0.648	0.710	0.837
Retrieval Relevance	0.860	0.796	0.619
Coherence	0.912	0.909	0.993
String Similarity	0.173	0.193	0.196
BLEU	0.0218	0.0224	0.0245
ROUGE-L	0.0830	0.1091	0.1118
Non-LLM Context Precision	0.0102	0.0167	0.0158
LLM Context Recall	0.213	0.198	0.334
LLM Context Precision	0.525	0.506	0.490

---

Project Structure

PaperRAG >
rag/
├── app.py                  # FastAPI entrypoint – exposes the RAG pipeline as API endpoints
├── requirements.txt        # All Python dependencies for the project
├── Dockerfile              # Docker setup for production-ready deployment
├── datasets.json           # Dataset / paper metadata for retrieval
├── Evaluate.ipynb          # Jupyter notebook for RAG evaluation and experimentation
├── Evaluate_ragas.py       # Script for evaluation using RAGAS metrics

├── Src/                    # Core application source code (agents, graph, prompts, utils)
│   ├── __init__.py
│   ├── agents.py           # Assembled LangGraph / LangChain agents (router, query generator, generator)
│   ├── prompts/            # All prompt templates for various agent components
│   │   ├── agent.txt
│   │   ├── generate_queries.txt
│   │   ├── generator.txt
│   │   ├── query_router.txt
│   │   ├── simple_generator.txt
│   │   └── step_back.txt
│   └── utils/              # Utility modules for state, nodes, and graph tooling
│       ├── nodes.py        # LangGraph node functions
│       ├── state.py        # RAG pipeline state management (TypedDict / pydantic)
│       ├── tools.py        # Tools used by agents (retrievers, API calls)
│       └── __init__.py

├── Utils/                  # Additional helper modules outside main source
│   ├── pdf_utils.py        # Extract text/metadata from PDFs via PyMuPDF
│   └── __pycache__/       

└── __pycache__/            

---