PaperPro - AI Powered Research Papers Query Engine

The Problem

Research papers are everywhere, but finding answers across multiple papers is painful. You either:

• Spend hours manually reading through PDFs (slow)
• Use keyword search and get irrelevant results (noisy)
• Copy paste text into ChatGPT and lose source attribution (dangerous)

PaperPro solves this. It understands research context semantically, finds the most relevant papers, and generates answers grounded in real citations.

---

Features

• Semantic querying over 20+ research papers

• Hybrid Retrieval:

  • BM25 keyword retrieval
  • Dense vector similarity retrieval

• Cohere reranking for relevance optimization

• Query rewriting using LLMs

• Citation-aware answer generation

• ChromaDB vector database with HNSW indexing

• Gemini 2.5 Flash integration

• DeepEval for RAG evaluation

• Streamlit frontend for interactive querying

• Optimized low latency inference pipeline

• Source aware contextual responses

---

System Architecture

User Query
     ↓
Query Rewriting
     ↓
Hybrid Retrieval
 ├── BM25 Retrieval
 └── Dense Vector Retrieval
     ↓
Document Deduplication
     ↓
Cohere Reranking
     ↓
Context Construction
     ↓
Gemini 2.5 Flash
     ↓
Grounded Answer + Citations

---

Tech Stack

Backend

• Python
• LangChain

Retrieval

• ChromaDB
• Sentence Transformers
• BM25 Retriever

LLMs

• Gemini 2.5 Flash
• Cohere Rerank API

Frontend

• Streamlit

Evaluation

• DeepEval

---

Core Components

1. Hybrid Retrieval

PaperPro uses a hybrid retrieval pipeline combining:

BM25 Retrieval

Captures:

• exact keyword matches
• acronyms
• paper-specific terminology
• lexical similarity

Dense Retrieval

Uses Sentence Transformers embeddings to capture:

• semantic meaning
• paraphrased queries
• contextual similarity

This improves retrieval quality significantly over standalone retrieval methods.

---

2. ChromaDB Vector Store

Dense embeddings are stored inside ChromaDB using:

• persistent vector storage
• HNSW indexing
• efficient similarity search

Benefits:

• fast retrieval
• scalable vector search
• persistent storage

---

3. Query Rewriting

Before retrieval, ambiguous or vague user queries are rewritten using Gemini.

Example:

Input:

How can models become more reliable?

Rewritten:

What techniques improve factual reliability in large language models during inference and deployment?

This improves:

• retrieval relevance
• semantic precision
• reranking quality

---

4. Cohere Reranking

Retrieved documents are reranked using Cohere Rerank API.

Purpose:

• reorder retrieved chunks by semantic relevance
• filter noisy retrievals
• improve context quality before generation

This significantly improves final answer quality.

---

5. Answer Generation

The final response is generated using:

• Gemini 2.5 Flash
• context-aware prompting
• source-grounded synthesis

The model:

• synthesizes across multiple papers
• cites relevant sources
• avoids hallucination when context is insufficient

---

Performance Optimization

Initial end-to-end query latency:

~2 minutes

Optimized latency:

~9 seconds

Optimizations Performed

Cached ChromaDB Loading

Used:

@st.cache_resource

to avoid reloading vector stores on every rerun.

Cached BM25 Retriever

Prevented reconstruction of BM25 indexes for each query.

Cached Metadata

Cached paper metadata and sidebar information to reduce Streamlit reruns.

Reduced Redundant LLM Calls

Optimized retrieval flow to minimize unnecessary generation overhead.

Persistent Model Initialization

Models are initialized once and reused across requests.

---

Evaluation Pipeline

PaperPro includes a DeepEval-based evaluation framework.

Metrics Evaluated

Answer Relevancy

Measures how well generated answers address the query.

Faithfulness

Measures whether generated responses remain grounded in retrieved context.

Contextual Recall

Measures whether the retrieval system fetched sufficient relevant information.

---

Example Query

Installation

1. Clone Repository

git clone https://github.com/yourusername/MLResearchRAG.git

cd MLResearchRAG

---

2. Create Virtual Environment

python -m venv venv

Activate:

Windows

venv\Scripts\activate

Linux/Mac

source venv/bin/activate

---

3. Install Dependencies

pip install -r requirements.txt

---

4. Configure Environment Variables

Create .env

GEMINI_API_KEY=your_key
COHERE_API_KEY=your_key

---

5. Run Application

streamlit run app.py

---

Future Improvements

• Multi-paper comparative reasoning
• Conversational memory
• PDF upload support
• Research graph visualization
• Agentic retrieval workflows
• Citation export
• Multi-modal paper understanding
• RAGAS evaluation integration
• Streaming token generation
• Redis caching layer
• GPU inference optimization

---

Deployment

The application is deployed using:

• Streamlit Cloud

Can also be deployed on:

• HuggingFace Spaces
• Render
• Railway
• AWS/GCP/Azure

---

Key Learnings

Through this project:

• implemented production-style RAG architectures
• optimized low-latency retrieval systems
• worked with vector databases and semantic search
• designed evaluation pipelines for generative AI systems
• improved retrieval grounding and hallucination reduction
• built scalable AI-assisted research workflows

---

License

MIT License

---

Author

Roopasree Computer Science Engineering, NIT Trichy

Made with <3 for Researchers