overview
The backend wraps the original biomedical pipeline when OpenRouter and ToolUniverse are available. In local recorded mode, it displays an original submission CSV trace with the real plan, tool calls, Tool Facts, and final answer.
Reasoning System
Plan-Act-Verify Biomedical Reasoning
A CURE-Bench Plan-Act-Verify system that uses a model planner, biomedical tools, curated Tool Facts, and a final answer pass.
Run demoGitHub placeholder
Published acc.
0.69564
fine-tuned GPT-4.1 + tools
Passes
2
plan then answer/verify
Tool families
6
FDA, DailyMed, RxNav, more
role
Portfolio integration: converted the original benchmark pipeline and submission artifacts into a readable question, plan, tool retrieval, verification, and answer display.
backend runner
Run original project workflow
Calls /api/biomedical/run on the FastAPI wrapper at http://localhost:8000.
question and choices
backend contract
Start the backend with uvicorn backend.main:app --reload --port 8000. The UI displays mode and provenance so recorded artifacts are clearly distinguished from live computation.
No backend result yet
Configure the inputs and run the backend to display the original project trace and outputs here.
architecture flow
Agent and model flow
The live pipeline trace appears in the backend runner after execution. This section shows the original project components that the backend wraps.
01
planner
GPT5Model.plan
Analyzes the stem and choices, extracts keywords, selects facts needed, and proposes biomedical tools.
02
retrieval
ToolAgent.collect
Runs curated biomedical tools and records success/failure traces for evidence gathering.
03
filter
Tool Fact Curator
Filters, deduplicates, clips, and diversifies successful facts before the answer pass.
04
verifier
Pass-2 Answer Prompt
Combines prior analysis, curated facts, and the full MCQ to produce one final answer letter.
tools and models
Components behind the demo
planner
GPT5Model.plan
Analyzes the stem and choices, extracts keywords, selects facts needed, and proposes biomedical tools.
retrieval
ToolAgent.collect
Runs curated biomedical tools and records success/failure traces for evidence gathering.
filter
Tool Fact Curator
Filters, deduplicates, clips, and diversifies successful facts before the answer pass.
verifier
Pass-2 Answer Prompt
Combines prior analysis, curated facts, and the full MCQ to produce one final answer letter.
example input
A pediatric generalized myasthenia gravis multiple-choice question with drug choices.
final result
The backend returns either a live original pipeline run or a provenance-backed recorded submission row from the original CURE-Bench outputs.
limitations
- Live mode requires OpenRouter credentials and ToolUniverse dependencies.
- Recorded mode is an original artifact but does not answer arbitrary new biomedical questions.
- This is benchmark reasoning output, not medical advice.