How snap-stanford/biomni works
Project type and purpose
• Biomni is an open-source Python library and CLI tool implementing a general-purpose biomedical AI agent (no dedicated frontend code in this repo).
• It orchestrates LLM calls, retrieval-augmented planning, domain-specific tools, and code execution to automate research tasks.
File structure
• Top-level:
• setup.py, setup.cfg, MANIFEST.in → packaging/build configuration
• biomni_env/ → conda/R environment and CLI tools installation scripts
• biomni/ → core Python package
• tutorials/, fig/, figs/ → usage examples and assets
• Inside biomni/:
• agent/ → agent controllers (A1, planning, react framework)
• model/ → retrieval logic (Vector store, embeddings)
• task/ → abstract Task definition + domain-specific tasks
• tool/ → a registry of domain tools, schemas, and support utilities
• biorxiv_scripts/ → data-prep scripts for generating new tasks from preprints
• utils.py, llm.py, env_desc.py → cross-cutting utilities and LLM wrappers
• Patterns:
• Plugin/registry pattern in tool_registry.py
• Modular package structure (agent – model – tool – task)
• Separation of concerns: planning vs. execution vs. retrieval
README for additional insights
• Confirms core components: agent.go(), environment setup, tool/database integrations
• Emphasis on extensibility: add new tools, datasets, benchmarks
• Mentions Web UI separately (hosted outside this repo)