MARVEL: Multi-Agent Research Validator & Enabler using LLMs

MARVEL v1.0.0 — Initial public release (accompanying MARVEL manuscript)

This release is the archival, citeable snapshot of the MARVEL codebase accompanying the manuscript:

"MARVEL: A Multi‑Agent Research Validator and Enabler using LLMs"

The MARVEL framework is a locally deployable, open-source system for domain-aware QA and assisted research with a fast path for simple queries and an optional DeepSearch mode for multi-step retrieval and synthesis.

A public demo is available at: https://ligogpt.mit.edu/marvel

What's included in this release

• marvel.py — Streamlit application (MARVEL-Standard + MARVEL-DeepSearch)
• config/ — configuration YAMLs (models, retrieval, data, prompts, etc.)
• scripts/ — setup + evaluation tooling
• Evaluation datasets: scripts/evaluation/datasets/
  • See: scripts/evaluation/datasets/README.md for file descriptions + schema

Quickstart (minimal)

1. Create conda environment (Linux/macOS/Windows):

• See README.md → "Conda environment setup"
• Helper: scripts/setup/setup_conda_env.py

1. Local inference (recommended): install Ollama + pull default models:

• See README.md → "Local LLM inference with Ollama"
• Defaults are defined in config/models.yaml

1. Select/build a FAISS vector store:

• Choose dataset via config/data.yaml (DocPDF, LatexData, TextData, JSONLData, or combined ALL_V2)
• FAISS stores persist under ./faiss/<DatasetName>/

1. Launch: streamlit run marvel.py

Optional services

• Groq (cloud LLM inference): set GROQ_API_KEY in config/retrieval.yaml
• Tavily (web search): set TAVILY_API_KEY in config/retrieval.yaml

Citation

• Use GitHub "Cite this repository" (CITATION.cff) or cite the arXiv paper.