How ur-whitelab/chemcrow-public works
Project type and purpose
• ChemCrow is an open-source Python package (library) that augments large-language models with chemistry-specific tools.
• It exposes a “ChemCrow” agent API (built on LangChain) that orchestrates calls to chemistry tool wrappers (RDKit, ChemSpace, reaction predictors, etc.), external APIs (OpenAI, SerpApi, RXN4Chem), and optional self-hosted microservices (Docker images for retrosynthesis and molecular-transformer).
• It also provides a lightweight Streamlit frontend for interactive use.
File structure
• Top-level:
• chemcrow/agents – orchestrator classes and prompt definitions
• chemcrow/tools – wrappers around chemistry libraries and external services
• chemcrow/data – static CSV datasets (e.g. chem_wep.csv)
• chemcrow/docker – Docker configurations for two self-hosted services (aizynthfinder and molecular-transformer)
• chemcrow/frontend – Streamlit callback handler and UI utilities
• tests/ – unit tests validating agents and tools
• dev-requirements.txt, setup.py – build and dependency definitions
• GitHub workflows for CI/CD
• The structure suggests a modular, plugin-style architecture: Agent layer → Tool layer → Data/external services.
README for additional insights
• The README emphasizes:
• Core functionality: instantiate ChemCrow(model, …).run(“task”)
• Dependencies on OpenAI API, SerpApi, RXN4Chem (or self-hosted alternatives)
• Optional microservices launched via Docker to replace hosted RXN4Chem endpoints
• A Streamlit demo on HuggingFace Spaces
Elements
• Blue for internal components (Agent, Tool Wrappers)
• Green for local libraries/data
• Orange for external APIs
• Purple for Docker microservices
• Grey for CI/CD