Back to Projects
🧠

Research Mind

LLM-Powered Research Copilot for Literature Review and Evidence-Grounded Question Answering

Status

Active Development

Timeline

6 months

Role

Full Stack Engineer

Problem Statement

Literature review is a time-consuming process that requires researchers to:

  • Search across multiple papers manually, spending hours finding relevant information
  • Extract and synthesize information from unstructured PDFs without tool support
  • Compare findings across papers to identify gaps and contradictions
  • Generate insights while maintaining evidence-grounding and avoiding hallucinations

Target Users: Graduate students, researchers, and academics conducting systematic literature reviews or rapid research synthesis.

Research Motivation

Why RAG?

Retrieval-Augmented Generation (RAG) combines dense retrieval with language models to ground answers in source documents, reducing hallucinations and enabling evidence-backed responses.

Why Hybrid Search?

Hybrid retrieval (BM25 + semantic embeddings) captures both lexical and semantic relevance, improving retrieval accuracy compared to single-mode search.

System Architecture

Backend Pipeline

Document Ingestion: Upload PDFs → Extract text with PyPDF

Chunking: Split text into overlapping chunks (512 tokens, 50 token overlap)

Embedding: Generate embeddings using HuggingFace sentence-transformers

Indexing: Store embeddings in FAISS with BM25 fallback index

Retrieval: Hybrid search combining semantic similarity + keyword matching

Generation: LLM processes retrieved documents + user query → grounded response

Frontend Interface

PDF Upload: Drag-and-drop interface for document management

Query Interface: Real-time question answering with retrieval transparency

Citation Display: Show source documents and retrieved chunks with highlights

Paper Comparison: Multi-document view for synthesizing findings

Technical Implementation

Backend Stack

  • Framework: FastAPI
  • Embeddings: HuggingFace sentence-transformers
  • Vector Store: FAISS
  • Retrieval: BM25 + semantic hybrid search
  • LLM: GPT-4 / Claude API
  • PDF Processing: PyPDF, LangChain

Frontend Stack

  • Framework: React with TypeScript
  • State Management: React Context
  • UI Components: TailwindCSS
  • API Integration: Axios
  • PDF Viewer: react-pdf
  • Deployment: Vercel

Technical Challenges & Solutions

Challenge 1: Handling Large PDFs

Problem: Large research papers (50+ pages) exceed token limits when chunked naively.

Solution: Implemented intelligent chunking with sliding window overlap, separating tables and figures to preserve structure while respecting LLM token limits.

Challenge 2: Hallucination Reduction

Problem: LLM generates plausible but unsupported answers when retrieval fails.

Solution: Implemented retrieval validation — LLM only generates answers if confidence score exceeds threshold; otherwise suggests retrieving more documents.

Challenge 3: Cross-Document Reasoning

Problem: Queries requiring synthesis across multiple papers often retrieve irrelevant sections.

Solution: Developed multi-query expansion strategy where LLM rephrases questions to retrieve diverse perspectives, improving recall for comparative analysis.

Methodology

Evaluation Framework

Evaluated system performance across three dimensions:

  • Retrieval Quality: NDCG@5, MRR — measuring relevance of retrieved documents
  • Generation Quality: ROUGE, BERTScore — comparing generated answers to gold references
  • Factuality: Manual annotation of hallucination rates and citation accuracy

Dataset & Benchmarking

Tested on 50 research papers from arXiv (NLP domain) with 200+ curated questions. Compared against keyword search baseline and single-embedding retrieval.

Results & Impact

Retrieval Accuracy

+45%

Hybrid search vs. semantic-only

Hallucination Reduction

-68%

With confidence-based filtering

User Satisfaction

8.2/10

From researcher feedback (n=15)

Key Findings

  • • Hybrid retrieval outperforms semantic-only and keyword-only search across all metrics
  • • Multi-query expansion improves cross-document reasoning by 31% for comparative questions
  • • Confidence-based filtering reduces hallucinations while maintaining answer quality
  • • Users prefer cited answers with source transparency over unsourced summaries

Future Work

Multi-agent orchestration for automated literature synthesis and research report generation

Structured extraction of claims, methodologies, and results for meta-analysis workflows

Interactive visualization of citation networks and research gaps across document collections

Integration with reference management systems (Zotero, Mendeley) for seamless workflow

Fine-tuned retrievers for domain-specific papers (biomedical, physics, computer science)

Hallucination evaluation framework with automated factuality scoring

Links & Resources

Built with Next.js, TypeScript, and TailwindCSS.

© 2026 Shridipa Dhar