Function calling with a custom LLM

With a custom LLM, Tried does not execute tools for you. The Functions section of the agent builder is hidden for custom LLM agents, and tools defined on an LLM or flow are not sent to your server. Your server decides when to call a tool, calls it, and decides what the agent says.

Tried gives you three things to work with:

1. Built-in call actions on a response or agent_interrupt frame: end_call, transfer_number (with show_transferee_as_caller) and digit_to_press.
2. Tool call events (tool_call_invocation and tool_call_result) that record your server's tool calls in the call transcript.
3. agent_interrupt to speak when a slow tool finishes, without waiting for the caller.

Built-in call actions#

These replace the end_call, transfer_call and press_digit tools that dashboard agents use.

Example: confirm and transfer in one response.

{
  "response_type": "response",
  "response_id": 7,
  "content": "Let me put you through to accounts.",
  "content_complete": true,
  "transfer_number": "+61290000000"
}

Only cold transfers by number are available from a custom LLM. Warm transfers, transfer agents, SMS and the other dashboard tool types are not reachable through the protocol. If you need them, run that logic in your own telephony integration or use a dashboard agent.

Running your own tools#

A typical turn with a tool looks like this:

{ "response_type": "response", "response_id": 4, "content": "One moment while I check.", "content_complete": false }
{ "response_type": "tool_call_invocation", "tool_call_id": "call_lookup_1", "name": "lookup_booking", "arguments": "{\"phone\":\"+61411111111\"}" }
{ "response_type": "tool_call_result", "tool_call_id": "call_lookup_1", "content": "{\"date\":\"Tuesday\",\"time\":\"14:00\"}" }
{ "response_type": "response", "response_id": 4, "content": " You're booked for Tuesday at 2pm.", "content_complete": true }

Keep the gap between response frames under 30 seconds. If no frame of a requested response arrives for 30 seconds, the call ends with error_llm_websocket_runtime.

What the tool events do#

tool_call_invocation and tool_call_result are recorded as transcript items on the call, in transcript_with_tool_calls:

[
  { "role": "user", "content": "When is my appointment?" },
  { "role": "tool_call_invocation", "tool_call_id": "call_lookup_1", "name": "lookup_booking", "arguments": "{\"phone\":\"+61411111111\"}" },
  { "role": "tool_call_result", "tool_call_id": "call_lookup_1", "content": "{\"date\":\"Tuesday\",\"time\":\"14:00\"}" },
  { "role": "agent", "content": "One moment while I check. You're booked for Tuesday at 2pm." }
]

They show in the call's transcript in Call History and are part of the call object returned by the API. They do not change what the agent says, and they are not sent back to your server in later requests (the worker sends transcript only, with agent and user items).

Send them only if you want the record. A server that never sends them works the same way.

Slow tools and agent_interrupt#

If a tool takes longer than a sentence or two, finish the current response with a holding line, then speak the result with agent_interrupt when it is ready:

{ "response_type": "response", "response_id": 9, "content": "I'm checking the warehouse system, this can take a moment.", "content_complete": true }

Later:

{
  "response_type": "agent_interrupt",
  "interrupt_id": 1,
  "content": "Thanks for waiting. Your order left the warehouse this morning.",
  "content_complete": true
}

agent_interrupt text is spoken as soon as it is complete, and its end_call, transfer_number and digit_to_press flags are applied after it. Use a new interrupt_id for each interrupt.

If the caller speaks while you wait, a new response_required arrives. Decide whether to answer it, or to answer it and still send the interrupt later.

Using call details in tools#

Turn on call_details in your config frame to receive the call object at connect time. Use from_number, to_number, metadata, retell_llm_dynamic_variables and collected_dynamic_variables as tool inputs (for example, look up the caller by from_number). See WebSocket server.

Gotchas#

• No tool definitions are sent to you. Define tools in your own model call.
• Every call action is applied after speech. If you need to hang up without saying anything, send content: "" with content_complete: true and end_call: true.
• transfer_number takes priority. If a response sets both transfer_number and end_call, the call is transferred, not ended.
• Stale ids are dropped. If the caller interrupts while your tool runs, frames for the old response_id are discarded, including the final answer. Answer the newest request, or use agent_interrupt for results that should be spoken anyway.
• metadata frames are not stored. Use tool events or your own database to keep per-call data.