Health Trifecta

Multi-agent CrewAI crew that simulates a small care team: a primary care physician, a registered dietitian, a certified personal trainer, and an HTML writer. Together they produce a structured health advisory narrative and a final HTML report grounded in a shared patient knowledge file (crew-level RAG).

Disclaimer

This project is for informational and educational use only. It is not medical advice, diagnosis, or treatment, and it is not a substitute for a licensed clinician, dietitian, or trainer.

Outputs are AI-generated and may be incomplete, outdated, or incorrect (including “hallucinations”). The sample patient file is illustrative; do not use this tool for real patient decisions without human professional oversight and appropriate safeguards (including privacy and consent).

What it does

1. Diagnosis and plan — The primary care physician task uses patient context from the knowledge base to draft a diagnosis and treatment-style plan.
2. Nutrition — The dietitian reviews that plan and adds evidence-based nutrition guidance, including a weekly Kosher-oriented nutrition plan.
3. Exercise — The trainer reviews the same clinical plan and adds a weekly exercise plan.
4. Integrated advisory — The physician synthesizes dietitian and trainer outputs into a comprehensive advisory report (including BMI-related framing and current vs target status in tabular form).
5. HTML deliverable — The HTML writer turns the advisory into a styled HTML document.

Orchestration is sequential (Process.sequential in src/health_trifecta/crew.py). Tasks declare context dependencies in src/health_trifecta/config/tasks.yaml, so downstream steps automatically receive upstream task outputs as input.

Agents and personas

| Agent | Role (summary) | Default LLM | |--------|----------------|-------------| | primary_care_physician | Holistic primary care framing; uses knowledge for patient facts; synthesizes specialist inputs in the advisory step. | openai/gpt-4o-mini | | registered_dietitian | Evidence-based nutrition advice to support the clinical plan. | openai/gpt-4o-mini | | certified_personal_trainer | Evidence-based exercise advice aligned with the clinical plan. | openai/gpt-4o-mini | | html_writer | Produces the final patient-facing HTML report from the advisory content. | anthropic/claude-opus-4-5-20251101 (override in agents.yaml if your API catalog differs) |

Full prompts (role, goal, backstory) live in src/health_trifecta/config/agents.yaml.

Knowledge sources and RAG

• Source: TextFileKnowledgeSource in src/health_trifecta/crew.py, wired as knowledge_sources on the Crew.
• Corpus: knowledge/patient-info.txt (example structured vitals and demographics). Paths are resolved relative to CrewAI’s knowledge/ convention from the configured file path in crew.py.
• Behavior: At run time, CrewAI indexes/embeds that content and retrieves relevant chunks when agents execute tasks (query-based RAG), so the physician step can ground answers in the file instead of inventing patient details.

To change the corpus, edit or replace knowledge/patient-info.txt and keep the TextFileKnowledgeSource file_paths in crew.py consistent with your layout.

Outputs

• output/health_advisory_report.html — Written by the html_writer_task (tasks.yaml output_file).

Installation

Requires Python >=3.10, <3.14. Dependencies are managed with uv.

pip install uv
cd health-trifecta
uv sync

The project pins crewai[google-genai,tools,anthropic] so OpenAI-, Google-, and Anthropic-routed models used in YAML are installable from one lockfile.

Environment

Create a .env in the project root with at least the keys below:

• OPENAI_API_KEY — Required for agents on OpenAI models.
• ANTHROPIC_API_KEY — Required for the HTML writer when using an anthropic/... model.
• MODEL (optional) — Default model hint for tooling; agents still follow per-agent llm in agents.yaml.
• CREWAI_STORAGE_DIR (optional) — Overrides CrewAI app data naming for SQLite and related storage; use a writable directory if you hit database permission errors.

Running

From the repository root:

crewai run

or:

uv run run_crew

Customizing

• src/health_trifecta/config/agents.yaml — Roles, goals, backstories, and llm strings per agent.
• src/health_trifecta/config/tasks.yaml — Task descriptions, context lists, expected_output, and output_file.
• src/health_trifecta/crew.py — Crew process, verbosity, and knowledge source definitions.
• src/health_trifecta/main.py — Kickoff inputs and CLI entrypoints (run, train, replay, test).

Support

• CrewAI documentation
• CrewAI GitHub
• Discord