Multiagent Paper Review System

A research system for automated scientific paper review using multi-agent architectures with Large Language Models (LLMs). The system processes academic papers and generates structured multi-perspective reviews through coordinated specialized agents.

Architecture

The system consists of three main components orchestrated via Docker Compose:

1. Papers Server (MCP Server)

A FastMCP-based Model Context Protocol server that provides paper search and retrieval capabilities.

• papersMCP.py: MCP server implementation exposing two tools:

  • search_articles(query): Semantic search using FAISS vector store, returns top-3 relevant paper chunks with relevance scores
  • get_article_content(id): Retrieves a configurable window (default 20KB) of text centered on the matched chunk position, enabling retrieval of longer context while avoiding loading entire large documents

• CreateVectorStore.ipynb: Jupyter notebook for batch processing incoming PDF papers:

  • Loads PDFs using PyMuPDF4LLM
  • Cleans and preprocesses text (removes tables, links, references, multi-line breaks)
  • Chunks documents with 10% overlap using RecursiveCharacterTextSplitter
  • Stores in FAISS vector database with metadata (title, area, doc_id, chunk position)
  • Saves processed text as markdown in vector_db/documents/

Papers are organized in papers-server/papers/ by research area subdirectories (CS, biology, nutrition).

2. Paper Export Client (Multi-Agent Reviewer)

A LangGraph-based multiagent system generating structured paper reviews from three distinct perspectives.

• ReviewerAgents.ipynb: Main review orchestration notebook implementing a StateGraph with four node types:
  • nice_reviewer: Generates favorable reviews emphasizing positive aspects
  • strict_reviewer: Critical review focusing on research gaps and weaknesses
  • neutral_reviewer: Balanced perspective avoiding strong opinions
  • editor: Synthesizes final review from all three perspectives

The output schema (Review model) includes:

• area: Main research area
• problem: Problem statement addressed
• step_by_step: Methodological approach
• conclusion: Process conclusion
• overall_review: Comprehensive markdown review (in Portuguese)

Each reviewer agent has access to the MCP server tools for cross-referencing other papers in the database.

3. Llama.cpp (configurable)

CUDA-enabled Docker container serving GGUF-format LLM models via OpenAI-compatible API:

• Chat model: NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf (4B parameters)
• Embedding model: embeddinggemma-300m-qat-Q4_0.gguf

Shared Utilities

common/utils.py: Core utilities for both server and client:

• LLM instantiation via lm-config.yaml (supports OpenAI-compatible, Google GenAI, Ollama providers)
• PDF loading via PyMuPDF4LLM
• Text cleaning functions (markdown tables, links, references, multi-line breaks)
• Document chunking with configurable size and overlap

Configuration

lm-config.yaml: Defines LLM provider settings. Can be altered to run cloud model:

main_language_model:
  provider: openai
  model: NVIDIA-Nemotron-3-Nano-4B-GGUF
  base_url: http://llamacpp:8082/v1
  
main_embedding_model:
  provider: openai
  model: embeddinggemma-300m-qat-GGUF
  base_url: http://llamacpp:8082/v1

Deployment

docker compose up --build

The client Jupyter notebook is accessible at http://localhost:8888.