Skip to main content

Prompts

Prompts in LlamaFarm are simple but powerful: you define static instructions in llamafarm.yaml, and the runtime merges them with chat history and (optionally) RAG context. This section outlines current capabilities and roadmap plans.

Prompt Configuration​

prompts:
  - name: default
    messages:
      - role: system
        content: >-
          You are a regulatory assistant. Provide concise answers and cite sources by title.
      - role: user
        content: "Use bullet points by default."
  • Prompts are named sets that can be selectively applied to models.
  • Messages within each prompt set are preserved in order and prepended to conversations.
  • Roles should match what your provider understands (system, user, assistant).
  • Models can specify which prompt sets to use via prompts: [list of names]; if omitted, all prompts stack in definition order.
  • Combine with RAG by including instructions explaining how to use context snippets (the server injects them automatically).

Best Practices​

  • Explain context usage: remind the model that context chunks contain citations or metadata.
  • Handle non-RAG scenarios: mention what to do when no documents are retrieved (“answer from general knowledge” or “state that no information was found”).
  • Keep prompts concise: long system instructions can reduce available tokens on smaller models.
  • Avoid conflicting instructions: align prompts with agent handler expectations (structured vs. simple chat).

Roadmap & Limitations​

  • Prompt templates, versioning, and evaluation tooling are in development. Track progress in the roadmap.
  • For now, dynamic templating (Jinja, variables) is not built-in—generate prompts upstream if needed.