Tool performance optimization #93
Conversation
Major simplification of the MCP tool interface to improve usability and maintainability. All domain tools now use consistent action parameters ("search", "get", "predict") instead of separate _searcher/_getter functions. This reduces cognitive overhead for users while following industry-standard patterns from projects like AIVA.
Breaking Changes:
- article_searcher/getter → article(action="search/get")
- trial_searcher/getter → trial(action="search/get")
- variant_searcher/getter → variant(action="search/get")
- gene/drug/disease_getter → gene/drug/disease(action="get")
- nci_*_searcher/getter → nci(resource="...", action="search/get")
- openfda_*_searcher/getter → fda(domain="...", action="search/get")
See CHANGELOG.md for complete migration guide.
Corrected parameter names and action patterns: - trial: Use detail parameter instead of separate get_* actions - variant: Use variant_id instead of id - alphagenome: Add action="predict" parameter - gene/disease/drug: Use id instead of *_id_or_name/symbol - fda: Fix parameter names (drug vs brand_name, etc.) and add more examples
|
Thanks @tkmamidi for the PR. Definitely agree that BioMCP consumes a lot of tokens if you turn on all of the tools at once. Any way you can selectively choose tools? I considered the "action" parameter too, but it has downsides that might confuse the LLM on what's available. We are going to take a hard look at both your solution and other alternatives with the goal of trimming down the context used. |
|
Thanks for the comment @imaurer ! |
|
Evals are a hard problem @tkmamidi ! One popular approach is LLM as a Judge, which I did for my "Deep Research" example: You could use an LLM as a Judge to compare your BioMCP-enabled program against a strong model with basic web search to see who comes up with "better answers". But the better thing would be to ask an expert. If you want to contribute to the "BioMCP-examples" Repo, I'd love that! |
This pull request introduces a major performance change to the BioMCP project by consolidating and simplifying the MCP tool interface, reducing the number of tools from 35 to 12. This reduces the agent token load from the tool docstrings and improves faster initialization. All domain tools now follow a consistent action-based pattern, which streamlines usage and improves maintainability. Extensive documentation and migration guides are provided to help users transition to the new interface, and all related code, guides, and examples have been updated accordingly.
Tool Interface Consolidation
action="search"oraction="get") for all domain tools, includingarticle,trial,variant,alphagenome,gene,drug,disease,fda, andnci.Documentation and Tutorial Updates
Code and Test Improvements
src/biomcp/tools/and improved parameter validation across all tools.Bug Fixes
hgvsparameter for general HGVS notation.sectionsarray parameters.