Adapters

LLM adapter for normalizing function and message formats.

This module provides the LLMAdapter class that normalizes interactions between the flow manager and Pipecat’s universal LLMContext. It handles:

  • Function name extraction from FlowsFunctionSchema

  • Function formatting into ToolsSchema for LLMSetToolsFrame

  • Summary message formatting

  • Summary generation via out-of-band LLM inference

class pipecat_flows.adapters.LLMAdapter[source]

Bases: object

Adapter for normalizing function and message formats.

Normalizes interactions between the flow manager and Pipecat’s universal LLMContext. Functions must be provided as FlowsFunctionSchema or FunctionSchema objects.

format_functions(functions: list[FunctionSchema | FlowsFunctionSchema]) → ToolsSchema | NotGiven[source]

Format functions into a ToolsSchema for use in LLMSetToolsFrame.

Parameters:

functions – List of function definitions (schema objects).

Returns:

ToolsSchema containing the functions, or NOT_GIVEN if no functions.

format_summary_message(summary: str) → dict[source]

Format a summary as a developer message.

Summary messages use the LLMContextMessage format (OpenAI-style), as summarization triggers an LLMMessagesUpdateFrame.

Parameters:

summary – The generated summary text.

Returns:

A developer message containing the summary.

async generate_summary(llm: Any, summary_prompt: str, context: LLMContext) → str | None[source]

Generate a summary by running a direct one-shot, out-of-band inference with the LLM.

Parameters:

  • llm – LLM service instance containing client/credentials.

  • summary_prompt – Prompt text to guide summary generation.

  • context – Context object containing conversation history for the summary.

Returns:

Generated summary text, or None if generation fails.