Paper Retrieval

This repository provides a paper retrieval and analysis system with three completed phases:

1. Plan 1: MCP infrastructure decoupling.
2. Plan 2: ReAct agent with dynamic tool routing.
3. Plan 3: Multi-agent architecture (Researcher -> Vision Expert -> Writer).

Current Status

All 3 plans are completed and integrated.

Environment Setup

This project uses conda environment pr.

conda activate pr
python -m pip install -r requirements.txt

If needed, use a fixed interpreter path:

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --help

Architecture Overview

Plan 1: MCP Server

• MCP server entry: mcp_server.py
• Server name: PaperBrain
• Exposed MCP tools include:
  • search_arxiv_tool
  • search_dblp_tool
  • download_and_extract_captions_tool
  • crop_figure_tool

Run MCP server:

C:/Users/lenovo/anaconda3/envs/pr/python.exe mcp_server.py

Plan 2: ReAct Dynamic Routing

• Runtime core: agent_runner.py
• Main entry: main.py
• ReAct builds on create_react_agent with ALL_TOOLS
• Streaming logs include Thought / Action / Observation / Final Answer

Plan 3: Multi-Agent Workflow

• State: workflow/multi_agent_state.py
• Nodes: workflow/multi_agent_nodes.py
• Graph: workflow/multi_agent_graph.py
• Runtime handoff logs:
  • [Researcher Agent working...]
  • [Vision Expert Agent working...]
  • [Writer Agent working...]

Run Modes

Main program entry is main.py.

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py [ARGS]

Mode 1: ReAct (default)

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --query "Search arXiv for recent vision-language model papers from the last 3 days"

Mode 2: Multi-Agent

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --mode multi-agent --query "Search arXiv for Vision-Language Models and Hallucination papers from last 3 days, then create a concise report"

Local vs Remote Retrieval Policy

The current behavior is:

1. Local intent is local-only.
   • For local queries, remote retrieval tools are disabled in ReAct mode.
   • In multi-agent mode, local intent processes local PDFs directly.
2. Remote retrieval is used when query intent is arXiv/DBLP/online retrieval.
3. Both modes still support report generation after analysis.

Local query example:

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --query "Analyze local PDFs under paper folder about Vision-Language Models and training-free, then generate one markdown report named react_local_training_free.md"

Remote query examples:

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --query "Search arXiv for recent Vision-Language Model papers from last 3 days, download the first one, analyze it, and generate one markdown report named arxiv_mode_report.md"

C:/Users/lenovo/anaconda3/envs/pr/python.exe main.py --query "Find CVPR 2025 papers from DBLP, download the first available PDF, analyze it with keywords Vision-Language Models and Hallucination, and generate one markdown report named dblp_mode_report.md"

Outputs

• Markdown reports are written to output/
• Extracted/cropped figures are written to output/images/

Validation

Run unit tests:

C:/Users/lenovo/anaconda3/envs/pr/python.exe -m unittest discover -s tests -v

Recent validation confirms tests pass in conda environment pr.