Citation Retrieval System

A multi-baseline citation retrieval system that combines BM25, dense retrieval (E5 and SPECTER), and LLM-based reranking to find relevant academic papers for citation contexts. Built with LangGraph for orchestration and designed for the ScholarCopilot dataset.

Overview

This system provides a comprehensive solution for academic citation retrieval, combining multiple retrieval strategies to find the most relevant papers for a given query or citation context. It uses a LangGraph-based workflow to orchestrate query reformulation, parallel retrieval from multiple sources, result aggregation, and LLM-based reranking.

The project consists of:

1. Server (Python): The core retrieval pipeline, evaluation scripts, and REST API
2. Client (Next.js): Optional web interface for finding citations (not required for evaluation)

Getting Started

For detailed setup instructions, evaluation guide, and API documentation, see server/README.md

The server README contains:

• Complete installation instructions (local setup or Docker)
• Evaluation CLI usage and examples
• API server documentation
• Configuration options
• Performance benchmarks
• Troubleshooting guide

Quick Links

• Evaluation & Setup: server/README.md - Start here for running evaluations on the ScholarCopilot corpus
• Frontend UI (optional): client/README.md - Web interface for interactive citation finding

Key Features

• Multi-Stage Retrieval Pipeline: Combines sparse (BM25) and dense (E5, SPECTER) retrieval methods
• Query Reformulation: Automatically expands and reformulates queries for better retrieval
• LLM-Based Reranking: Uses language models (Ollama or OpenAI) to rerank results
• Reciprocal Rank Fusion: Intelligently combines results from multiple retrievers
• ScholarCopilot Integration: Built-in support for the ScholarCopilot dataset format
• GPU-Accelerated: Batch processing for efficient dense retrieval
• Comprehensive Evaluation: Compare individual retrievers vs full pipeline with detailed metrics

Architecture Overview

The system uses a LangGraph-based workflow with the following stages:

Query Input
    ↓
Query Reformulator (expands queries)
    ↓
    ├──→ BM25 Retriever (sparse retrieval)
    ├──→ E5 Retriever (dense retrieval)
    └──→ SPECTER Retriever (academic paper embeddings)
    ↓
Aggregator (combines and fuses results)
    ↓
Reranker (LLM-based reranking)
    ↓
DSPy Picker (optional, citation selection)
    ↓
Final Ranked Results

For detailed information about each component, see server/README.md.