Struct LlmExtractor 

pub struct LlmExtractor { /* private fields */ }

LLM-backed turn extractor that sends transcript windows to an OOB LLM with a structured extraction prompt.

Implementations

impl LlmExtractor

pub fn new( name: impl Into<String>, llm: Arc<dyn BaseLlm>, prompt: impl Into<String>, window_size: usize, ) -> LlmExtractor

Create a new LLM-backed extractor.

• name: key for storing results in State
• llm: the out-of-band LLM to use for extraction
• prompt: system instruction describing what to extract
• window_size: how many recent turns to include

pub fn with_thinking_budget(self, budget: Option<u32>) -> LlmExtractor

Set the thinking budget sent with each extraction request.

The default is Some(0): extraction reads a few turns, and the turn pipeline waits for it, so thinking costs seconds of latency on every turn. Measured on gemini-flash-latest, a budget of 0 cut a short extraction from 7.7 s to 2.3 s with the same results. None leaves the model’s default. When a model rejects a budget of 0 (some only work in thinking mode, and gemini-3.5-flash-lite needs at least 1), the extractor retries with FALLBACK_THINKING_BUDGET, then without a budget, and keeps sending the first one the model accepted.

pub fn with_min_words(self, n: usize) -> LlmExtractor

Set the minimum word count in the last user utterance to trigger extraction.

Turns where the user said fewer than n words will skip the LLM call. Useful for filtering out “uh huh”, “ok”, “yes” style responses.

pub fn with_schema(self, schema: Value) -> LlmExtractor

Set a JSON Schema for structured output.

When set, the schema is included in the prompt to guide the LLM toward producing valid JSON matching the schema.

pub fn with_trigger(self, trigger: ExtractionTrigger) -> LlmExtractor

Set the trigger mode for this extractor.

pub fn with_promotions(self, rules: Vec<FieldPromotion>) -> LlmExtractor

Set explicit field promotion rules.

Once promotion rules are present, top-level fields are no longer automatically flattened into state; only accepted rules promote.

Trait Implementations

impl TurnExtractor for LlmExtractor

fn extract_with_state<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, window: &'life1 [TranscriptTurn], state: &'life2 State, ) -> Pin<Box<dyn Future<Output = Result<Value, LlmError>> + Send + 'async_trait>>

where 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, LlmExtractor: 'async_trait,

Extract, telling the model which fields are already known so it returns only what the latest turn adds or changes.

fn name(&self) -> &str

Name of this extractor (used as the State key).

fn window_size(&self) -> usize

How many recent turns this extractor needs.

fn should_extract(&self, window: &[TranscriptTurn]) -> bool

Whether this extractor should run for the current turn.

fn trigger(&self) -> ExtractionTrigger

The trigger mode for this extractor.

fn promotion_rules(&self) -> &[FieldPromotion]

Field promotion rules for this extractor.

fn may_write(&self, key: &str) -> bool

Whether this extractor may write the state key key.

fn extract<'life0, 'life1, 'async_trait>( &'life0 self, window: &'life1 [TranscriptTurn], ) -> Pin<Box<dyn Future<Output = Result<Value, LlmError>> + Send + 'async_trait>>

where 'life0: 'async_trait, 'life1: 'async_trait, LlmExtractor: 'async_trait,

Extract structured data from the transcript window.

fn on_complete(&self) -> Option<OnComplete>

An optional agent to run when this extractor’s results land in state — the on_complete(dispatch(agent)) effect. Fired by the pipeline after promotion, only when the extractor produced a non-empty object.