DICE API Documentation/com.embabel.dice.common.resolver/LlmCandidateBakeoff

LlmCandidateBakeoff

data class LlmCandidateBakeoff(llmOptions: <Error class: unknown class>, ai: <Error class: unknown class>, promptStrategy: BakeoffPromptStrategy = BakeoffPromptStrategies.COMPACT) : CandidateBakeoff

Uses an LLM to select the best match from multiple candidates.

This is more efficient than evaluating candidates one-by-one, and allows the LLM to compare and contrast options. Uses structured output for reliable parsing.

Two built-in prompt strategies are available in BakeoffPromptStrategies:

BakeoffPromptStrategies.COMPACT: Minimal prompts (~100-200 tokens) for fast, cheap resolution. Does not include descriptions - may cause issues with similarly-named entities.
BakeoffPromptStrategies.FULL: Detailed prompts (~400-600 tokens) for complex disambiguation.

Custom strategies can be implemented for domain-specific prompt requirements.

Returns the best matching candidate, or null if none match.

Example usage (Java-friendly builder):

var bakeoff = LlmCandidateBakeoff
    .withLlm(llmOptions)
    .withAi(ai)
    .withPromptStrategy(BakeoffPromptStrategies.FULL);

Parameters

llmOptions

LLM configuration (model, temperature, etc.)

The Embabel AI instance for LLM calls

promptStrategy

Strategy for building prompts (default: COMPACT)

Constructors

LlmCandidateBakeoff

constructor(llmOptions: <Error class: unknown class>, ai: <Error class: unknown class>, promptStrategy: BakeoffPromptStrategy = BakeoffPromptStrategies.COMPACT)

Types

Companion

object Companion

Functions

selectBestMatch

open override fun selectBestMatch(suggested: SuggestedEntity, candidates: List<<Error class: unknown class><<Error class: unknown class>>>, sourceText: String?): <Error class: unknown class>?

Select the best matching candidate from a list.

withPromptStrategy

fun withPromptStrategy(strategy: BakeoffPromptStrategy): LlmCandidateBakeoff

Set the prompt strategy.