voicemate.agent.graph

The LangGraph agent: recall → agent ⇄ tools.

  • recall embeds the user message, loads matching facts, past exchanges and research, and builds the per-turn context block (about 30-50 ms, so no speculation is needed).

  • The system prompt is static and the context is attached only to the newest user message, so Ollama can reuse its KV cache for the long, tool-laden prompt prefix (see prompts.py).

  • agent calls the chat model with all tools bound. After max_tool_rounds tool rounds the tools are unbound so the model has to answer.

  • tools runs the requested tools; exceptions become error messages the model can read.

Conversation history is persisted per thread_id by the checkpointer. Events for the UI are sent through the emit callback in config["configurable"].

Attributes

Classes

AgentState

Graph state.

Functions

make_chat_model(config[, model]) → langchain_core.language_models.BaseChatModel

The Ollama chat model with thinking disabled (latency matters more for voice).

recent_history(messages[, limit]) → list

The last limit messages, starting at a user message.

with_context(messages, context) → list[langchain_core.messages.AnyMessage]

Copy of messages whose newest user message is prefixed with context.

warm_up(llm, tools, config) → None

Prefill the static prompt prefix so the first real turn hits Ollama's KV cache.

build_graph(llm, tools, config[, memory, ...]) → Any

Compile the agent graph.

Module Contents

voicemate.agent.graph.HISTORY_MESSAGES: int = 24
voicemate.agent.graph.ANSWER_NUDGE: str = "You have the tool results now: answer the user's question in full."
voicemate.agent.graph.NO_MORE_TOOLS: str = 'Tool budget for this turn is used up: answer now with what you have.'
class voicemate.agent.graph.AgentState[source]

Bases: TypedDict

Graph state.

Variables:
  • messages – Conversation history (appended through the add_messages reducer).

  • lang – Reply language of the current turn.

  • context – Per-turn context block (time, language, memories) for the newest message.

  • tool_rounds – Tool rounds used in the current turn.

Initialize self. See help(type(self)) for accurate signature.

messages: Annotated[list[langchain_core.messages.AnyMessage], langgraph.graph.message.add_messages]
lang: str
context: str
tool_rounds: int
voicemate.agent.graph.make_chat_model(config: voicemate.config.LLMConfig, model: str | None = None) → langchain_core.language_models.BaseChatModel[source]

The Ollama chat model with thinking disabled (latency matters more for voice).

Parameters:
  • config – LLM settings.

  • model – Model tag; defaults to the preferred profile’s model.

voicemate.agent.graph.recent_history(messages: collections.abc.Sequence[langchain_core.messages.AnyMessage], limit: int = HISTORY_MESSAGES) → list[source]

The last limit messages, starting at a user message.

Starting at a user message keeps tool results together with the call that produced them.

voicemate.agent.graph.with_context(messages: list[langchain_core.messages.AnyMessage], context: str) → list[langchain_core.messages.AnyMessage][source]

Copy of messages whose newest user message is prefixed with context.

async voicemate.agent.graph.warm_up(llm: langchain_core.language_models.BaseChatModel, tools: list[langchain_core.tools.BaseTool], config: voicemate.config.Config) → None[source]

Prefill the static prompt prefix so the first real turn hits Ollama’s KV cache.

voicemate.agent.graph.build_graph(llm: langchain_core.language_models.BaseChatModel, tools: list[langchain_core.tools.BaseTool], config: voicemate.config.Config, memory: voicemate.memory.store.MemoryStore | None = None, checkpointer: Any = None, now: collections.abc.Callable[[], datetime.datetime] = lambda : ...) → Any[source]

Compile the agent graph.

Parameters:
  • llm – Chat model supporting tool calling.

  • tools – Tools to bind.

  • config – Application config.

  • memory – Long-term memory used by the recall node (skipped when None).

  • checkpointer – LangGraph checkpointer for per-thread history.

  • now – Clock for the turn context.

Returns:

The compiled graph.