How katherlab/onprem-medical-agents works
On-Prem Medical Agents runs benchmark-configured, multi-turn doctor–patient simulations, using patient-case data and clinical tools to produce conversation logs; separate runners evaluate diagnoses and merge runs, while a confidence pipeline derives features and visualizations. The graph preserves MIMIC and VivaBench paths and the evaluation and analysis branches. Unexcerpted capabilities and wiring are included only where README or source-index references support them; some internal runtime routing remains uncertain.
Main entry points: simulate.py Simulation runner · executor_agent.py Doctor and patient agents · vis_confidence.py Confidence pipeline · conv.py Dialogue engine
Components 21
Simulation
- Agent settings
src/configs/agent_config.py - Simulation runner
src/simulate.py - Dialogue engine
src/conv.py - Doctor and patient agents
src/multi_agents/executor_agent.py - Benchmark prompts
src/prompts.py - Conversation logs
src/mes_collect.py
Clinical data and tools
- MIMIC case loader
src/dataset/mimic_dataset.py - VivaBench case loader
src/dataset/vivabench_dataset.py - Clinical tool definitions
src/tools/tool_request.py - Tool dispatch
src/tools/tool_handler.py - Clinical tool execution
src/tools/tool_execs.py - VivaBench tools
src/tools/tool_vivabench.py - Procedure search index
src/qdrant_collection.py
Evaluation
- Diagnosis evaluator
src/evaluate.py - Evaluation methods
src/eval/evaluators.py - Run aggregation
src/merge_runs.py
Confidence analysis
- Confidence pipeline
src/vis_confidence.py - Analysis data loader
src/confidence/data_loader.py - Feature engineering
src/confidence/feature_engineering.py - Confidence analysis
src/confidence/analysis.py
Other components
- Researcher
Connections 20
- Researcher → Simulation runner: starts simulation
README.md - Simulation runner → Agent settings: uses settings
README.md - Simulation runner → MIMIC case loader: loads MIMIC cases
README.md - Simulation runner → VivaBench case loader: loads VivaBench cases
README.md - Simulation runner → Dialogue engine: runs dialogue
src/simulate.py - Simulation runner → Doctor and patient agents: configures agents
src/simulate.py - Doctor and patient agents → Benchmark prompts: uses instructions
src/multi_agents/executor_agent.py - Dialogue engine → Conversation logs: records turns
src/conv.py - Doctor and patient agents → Clinical tool definitions: invokes tools
src/multi_agents/executor_agent.py - Clinical tool definitions → Tool dispatch: dispatches requests
src/tools/tool_handler.py - Tool dispatch → Clinical tool execution: fetches results
src/tools/tool_handler.py - Clinical tool execution → Procedure search index: searches procedures
src/tools/tool_execs.py - Doctor and patient agents → VivaBench tools: uses benchmark tools
- Researcher → Diagnosis evaluator: starts evaluation
README.md - Diagnosis evaluator → Evaluation methods: evaluates diagnoses
src/eval/evaluators.py - Researcher → Run aggregation: aggregates runs
README.md - Researcher → Confidence pipeline: starts analysis
README.md - Confidence pipeline → Analysis data loader: loads aggregates
README.md - Confidence pipeline → Feature engineering: computes features
README.md - Feature engineering → Confidence analysis: analyzes features