How yahskapar/personal-health-insights-agent works
PHIA is a research prototype that answers questions about wearable activity, sleep, and wellness using a ReAct LLM agent. The main workflow is to load a user’s summary and activity data, ask a question in the demo, and let the agent analyze the data with Python and optionally search the web before returning an answer. Few-shot notebooks guide the agent; evaluation artifacts and figure notebooks support the paper’s research results. The runtime links are grounded in sampled source; evaluation and figure wiring is only documented at a high level.
Main entry points: phia_agent.py ReAct agent · phia_demo.ipynb PHIA demo · auto_eval Evaluation results · data_utils.py Wearable data loader
Components 16
User Interaction
- PHIA demo
phia_demo.ipynb
Agent Runtime
- ReAct agent
phia_agent.py - Prompts and exemplars
prompt_templates.py - Python data tool
phia_agent.py - Web search tool
phia_agent.py - Wearable data loader
data_utils.py
Research Artifacts
- Few-shot examples
few_shots - Synthetic users
synthetic_wearable_users - Deidentified users
real_wearable_users - Query sets
Objective Query - PHIA.xlsx - Open-ended queries
Open-Ended Query - PHIA.xlsx - Evaluation results
data/auto_eval - Human annotations
data/human_label_experiments - Paper figures
figs
Other components
- Wearable-data user
- Tavily search
Connections 16
- Wearable-data user → PHIA demo: asks question
- PHIA demo → Wearable data loader: loads data
- Wearable data loader → Synthetic users: reads files
- Wearable data loader → Deidentified users: reads files
- PHIA demo → ReAct agent: invokes
- ReAct agent → Prompts and exemplars: sets prompt
- Prompts and exemplars → Few-shot examples: loads examples
- ReAct agent → Python data tool: dispatches code
- Python data tool → PHIA demo: analyzes data
- ReAct agent → Web search tool: dispatches search
- Web search tool → Tavily search: queries
- ReAct agent → PHIA demo: returns answer
- Query sets → Evaluation results: evaluated in
- Open-ended queries → Evaluation results: evaluated in
- Human annotations → Evaluation results: labels outputs
- Evaluation results → Paper figures: informs figures