Manage Prompts

All LLM prompts are stored in the database and versioned — never hardcoded in application code.

How Prompt Versioning Works

Each agent has a row in the agent_prompts table with:

• agent_name — identifies the agent (e.g., factuality_checker)
• version — integer version number
• model — the Claude model to use (e.g., claude-sonnet-4-6)
• system_prompt — the full system prompt text
• active — boolean flag; only one version per agent is active at a time

When an agent runs, call_agent() loads the active prompt for that agent from the database.

Updating a Prompt

Prompts are updated via Alembic data migrations. This ensures:

• Every prompt change is tracked in version control
• Changes can be rolled back
• Staging and production stay in sync via alembic upgrade head

To update a prompt:

1. Create a new migration (see Database Migrations)
2. In upgrade(): deactivate the old version, insert the new one
3. In downgrade(): delete the new version, reactivate the old one
4. Apply: uv run alembic upgrade head

Viewing Prompts in the Dashboard

The Prompts page in the Streamlit dashboard lets you:

• View all agent prompts and their versions
• See which version is currently active
• Compare prompt text across versions

Navigate to the Prompts page from the sidebar after logging in.

Prompt Conventions

• Use clear, structured system prompts with numbered steps
• Include output format specifications (JSON schema)
• For medical content, include safety language and hedging requirements
• Reference the agent's Pydantic output schema in the prompt so the model knows the expected structure