Skip to main content
Sends a prompt to an LLM for analysis, summarization, classification, or transformation. Display name “AI Call”. On by default for new agents. response is raw text, not a parsed object. If you asked for JSON it is usually wrapped in a ```json fence — strip the fence, then json.loads. Do not index into response as if it were already an object. When quoting extracted values in the prompt, convert {value, word_id_groups: [N]} to value[cite:N]. Never put word_id_groups in the prompt.

​
Authentication and enablement

No integration. Uses the agent’s LLM. Default-on.

​
Inputs

  • prompt (required): the per-call question / instruction.
  • system_prompt (optional).
  • cache_context (optional): stable content identical across calls (documents, reference data). Cached separately from prompt.
  • cache_input (optional bool).

​
Output

{tool_name, response, model, tokens_used, completed_at, metadata}.

​
Limits and side effects

  • Billed as an LLM call. Prefer passing prior tool output through generate_file content instead of re-asking the model to format a CSV.
  • No external writes.

​
Expected errors

  • Empty prompt.
  • Provider/model failure (retried internally where configured).