Multi-Agent Research Paper to Codebase Pipeline

Upload any research paper (PDF or arXiv ID) → AI agents analyze it → Generate a complete, runnable Python codebase → Download as ZIP

Transform arXiv papers into production-ready projects using a multi-agent workflow Analyst → Architect → Coder.

---

Table of Contents

• What It Does
• Generated Project Example
• Features
• Quick Start
  • 1. Clone & Install
  • 2. Configure LLM
  • 3. Run
• Usage Flow
  • Agent Roles
  • Key Design Decisions
• Project Structure
• Performance
• Contributing
• License
• Acknowledgments

---

What It Does

Most research papers are never implemented. Reading a paper, understanding its architecture, designing the project structure, and writing the code can take days — even for experienced engineers.

This pipeline does it automatically:

1. Reads the paper using RAG (FAISS vector search over paper chunks)
2. Analyses it, extracts components, equations, data flow, inputs/outputs
3. Plans a full Python project, files, classes, functions, algorithm steps
4. Codes every file — using the most relevant paper sections as context per file
5. Packages everything — folder structure, requirements.txt, README.md, downloadable ZIP

---

Generated Project Example

For "Attention Is All You Need" paper (arXiv:1706.03762):

output/projects/transformer-attention/
├── src/
│   ├── encoder.py         # MultiHeadAttention + FeedForward
│   ├── decoder.py         # DecoderLayer with encoder-decoder attention
│   ├── attention.py       # ScaledDotProductAttention
│   ├── model.py           # Transformer
│   └── main.py            # End-to-end demo
├── requirements.txt       # torch, numpy
└── README.md              # How to train/extend

---

Features

Feature	Description
Paper Ingestion	Upload local PDF or enter arXiv ID automatic text extraction and chunking
FAISS RAG	Vector index over paper chunks per-agent, per-file context retrieval
Multi-Agent Pipeline	Analyst → Architect → Coder three specialized agents, each with targeted prompts
Streamlit UI	Step-by-step flow: upload → review plan → generate code → download ZIP
Plan Review	Inspect every file, class, function, and algorithm step before generating code
Project Export	Full folder structure with src/, requirements.txt, README.md per project
Syntax Validation	ast.parse() on every generated file, broken files flagged in UI
Post-Processing	Auto-fix main_file prefix, auto-collect dependencies, phantom class removal
Debug Logs	Raw agent outputs saved per run for prompt tuning and debugging

---

🚀 Quick Start

1. Clone & Install

git clone https://github.com/Ravevx/multi-agent-paper-pipeline.git
cd multi-agent-paper-pipeline

python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

2. Configure LLM

Local LM Studio (default):

# Start LM Studio server at http://127.0.0.1:1234/v1
# Set your model in config.py
LMSTUDIO_URL=http://127.0.0.1:1234/v1
LMSTUDIO_MODEL=your-model-name

3. Run

streamlit run app.py

Usage Flow

1. Upload PDF or enter arXiv ID
2. [Analyse] → Build RAG → Analyst → Architect → Plan
3. Review plan (files, classes, functions, steps)
4. [Approve] → Generate code file-by-file
5. Download ZIP or use local copy

Agent Roles

Agent	Input	Output	Key Prompt Rule
Analyst	Paper chunks (RAG)	5-section technical analysis	Start with ## 1. no preamble, no hedging
Architect	Analysis text	ProjectPlan JSON	Class names must cross-reference, no phantom classes
Coder	File spec + RAG context	Python source file	EXACT NAMES box at top, enforces plan class names

Key Design Decisions

• RAG per file: Each file gets its own targeted RAG query using its class names and logic summary, so attention.py retrieves attention equation chunks, not encoder chunks
• Plan validation: _validate_plan() removes phantom classes (classes listed in main.py but not defined anywhere), preventing ImportError at runtime
• nn.Module enforcement: fix_nn_module() detects torch files and adds (torch.nn.Module) inheritance where missing
• No architect truncation fallback: If main.py algorithm steps are empty (truncated output), crew_runner.py rebuilds them from the actual class/function list

---

Project Structure

.
├── app.py                 # Streamlit app
├── crew_runner.py         # Main orchestration
├── crew_tasks.py          # Agent prompts
├── crew_agents.py         # Agent definitions
├── rag_store.py           # FAISS RAG
├── paper_tools.py         # PDF/arXiv extraction
├── project_planner.py     # Data models
├── config.py              # Settings
├── requirements.txt       # Dependencies
├── output/
│   ├── papers/            # Cached PDFs
│   └── projects/          # Generated codebases
│       └── <project_name>/
│           ├── src/
│           ├── requirements.txt
│           └── README.md
└── screenshots/           # UI screenshots

Performance

Paper	Files	Time	Syntax Errors
Attention Is All You Need	7	12 min	0
BERT	12	18 min	1
GPT-2	9	14 min	0

Project Architecture

                               
                           ┌───────────────────┐
                           │   RAG Index       │
                           └────────┬──────────┘
                                    │
                           ┌───────────────────┐
                           │Angent (Analyst)   │
                           │  (Paper Analysis) │
                           └────────┬──────────┘
                                    │
                           ┌───────────────────┐
                           │Angent (Architect) │
                           │  (Project Plan)   │─────┐
                           └────────┬──────────┘     |
                                    │                | 
                           ┌───────────────────┐     |
                           │ Human-in-the-loop │     |
                           │  Approve / Edit   │     |
                           └───────┬───────────┘     |
                                   │  Approved? No   |
                  ┌────────────────┴─────────────────┘
                  │                                  
           Approved? Yes                       
                  │                                                        
          ┌─────────────┐ 
          │Agent (Coder)│                  
          │ Generate    │                  
          │  Files      │                 
          └─────┬───────┘                 
                │                                 
                └───────────┐                                   
                   ┌─────────────────────────────┐
                   │ Generate Project Folder /    │
                   │ src/, requirements.txt,     │
                   │ README.md, ZIP Export       │
                   └─────────────────────────────┘

---

Contributing

Contributions are very welcome. Good areas to work on:

• Better prompts for specific domains (computer vision, reinforcement learning, graph networks)
• Alternative RAG backends (Chroma, Pinecone, Weaviate)
• Support for LaTeX source papers not on arXiv PDF
• Training loop generation extend Coder to also produce train scripts
• Unit tests for crew_runner.py post-processing functions

License

MIT License LICENSE

##Acknowledgments

• CrewAI — multi-agent orchestration
• FAISS — vector search
• Streamlit — amazing UI framework
• Research community — for the papers that power this!

---