BioThings / SmartAPI KPClient tool

[ahueb]

Summary

The Knowledge Provider (KP) tool provides unified access to BioThings Translator / SmartAPI-backed data sources through KPClient (biomni/tool/kp_tool.py).
This PR adds the KPClient tool so the agent can retrieve KG and provider-backed biomedical data (TRAPI + BioThings) through one interface.

Changes Made

Added biomni/tool/kp_tool.py containing the KPClient class and KP surface functions in biomni/tool/database.py and their descriptions in biomni/tool/tool_description/database.py:

• ensure_fresh
• list_kps
• describe_kp
• query_kp
• batch_query_kp
• scroll_kp

KP is selected when prompts request data from sources it covers (Reactome, STRING, DGIdb, DrugCentral, MyGene/MyVariant/MyChem/MyDisease, GO/HPO/Mondo/DOID/UBERON, ARAX/BTE/MolePro/SPOKE/COHD/mediKanren/RTX KG2).

Future improvements

Tool choice is LLM-mediated, a future improvement would be an execution-time guardrail that enforces at least one query_kp/batch_query_kp call whenever requested data is in KP-covered sources.

Testing

import sys

sys.path.append("../")

from biomni.agent import A1

agent = A1(llm="claude-sonnet-4-5-20250929")

# One prompt per data type
log = agent.go("please locate annotation data for gene ID 1017")
log = agent.go("please locate variant annotation data for rs7412")
log = agent.go("please locate pathway data for TP53")
log = agent.go("please locate protein interaction data for EGFR")
log = agent.go("please locate drug-target association data for imatinib")
log = agent.go("please locate disease-gene association data for Alzheimer's disease")
log = agent.go("please locate ontology mapping data for BRCA1")
log = agent.go("please locate multi-hop evidence connecting KRAS to colorectal cancer")

✅ Selection testing across the prompt set all passed