Crate brainwires_reasoning 

Brainwires Reasoning

Layer 3 — Intelligence. Provider-agnostic reasoning primitives for the Brainwires Agent Framework.

This crate owns:

• plan_parser — extract numbered task steps from LLM plan output.
• output_parser — parse structured data (JSON, regex) from raw LLM text.
• Local inference scorers — provider-agnostic TIER 1/2 components for routing, validation, complexity scoring, summarisation, retrieval gating, relevance ranking, strategy selection, and entity enhancement. All accept Arc<dyn Provider> and fall back to pattern-based logic when the provider is unavailable.

Scope note — prompting stays in brainwires-knowledge

The original architectural plan (sleepy-popping-falcon.md) called for prompting to move here too. It didn’t — prompting is tightly coupled to bks_pks inside brainwires-knowledge and a move would have pulled an entire knowledge-store dependency into this crate. The deviation is intentional; tests and consumers of prompting should continue to target brainwires_knowledge::prompting.

Configuration

Enable via LocalInferenceConfig:

[local_llm]
enabled = true
use_for_routing = true
use_for_validation = true
use_for_complexity = true
use_for_summarization = true

Re-exports

pub use output_parser::JsonListParser;

pub use output_parser::JsonOutputParser;

pub use output_parser::OutputParser;

pub use output_parser::RegexOutputParser;

pub use plan_parser::ParsedStep;

pub use plan_parser::parse_plan_steps;

pub use plan_parser::steps_to_tasks;

pub use strategies::ChainOfThoughtStrategy;

pub use strategies::ReActStrategy;

pub use strategies::ReasoningStrategy;

pub use strategies::ReflexionStrategy;

pub use strategies::StrategyPreset;

pub use strategies::StrategyStep;

pub use strategies::TreeOfThoughtsStrategy;

Modules

output_parser

Structured output parsers for LLM responses. Structured output parsing for LLM responses

plan_parser

Plan text parser for extracting steps from LLM output. Plan Parser - Extract tasks from plan content

strategies

Reasoning strategies: CoT, ReAct, Reflexion, Tree-of-Thoughts. Named Reasoning Strategies — ReAct, Reflexion, Chain of Thought, Tree of Thoughts

Structs

ClassificationResult

Result of classification with confidence

ComplexityResult

Result of complexity scoring

ComplexityScorer

Complexity scorer for task difficulty assessment

ComplexityScorerBuilder

Builder for ComplexityScorer

EnhancedEntity

An entity extracted by LLM

EnhancedRelationship

A semantic relationship between entities

EnhancementResult

Result of entity enhancement

EntityEnhancer

Entity enhancer using a provider

EntityEnhancerBuilder

Builder for EntityEnhancer

ExtractedFact

Key fact extracted from content

InferenceTimer

Measure inference latency.

LocalInferenceConfig

Configuration for local inference components.

LocalRouter

Router for semantic query classification

LocalRouterBuilder

Builder for LocalRouter

LocalSummarizer

Summarizer for context compression

LocalSummarizerBuilder

Builder for LocalSummarizer

LocalValidator

Validator for semantic response validation

LocalValidatorBuilder

Builder for LocalValidator

RelevanceResult

Result of relevance scoring

RelevanceScorer

Relevance scorer for context re-ranking

RelevanceScorerBuilder

Builder for RelevanceScorer

RetrievalClassifier

Retrieval classifier for enhanced gating

RetrievalClassifierBuilder

Builder for RetrievalClassifier

RouteResult

Result of route classification

StrategyResult

Result of strategy selection

StrategySelector

Strategy selector for MDAP decomposition

StrategySelectorBuilder

Builder for StrategySelector

SummarizationResult

Result of a summarization operation

Enums

FactCategory

Category of extracted facts

LocalRetrievalNeed

Result of retrieval classification

RecommendedStrategy

Recommended decomposition strategy

RelationType

Types of relationships we detect semantically

SemanticEntityType

Enhanced entity type with semantic classification

TaskType

Task type classification

ValidationResult

Result of local validation

Functions

log_inference

Log a local inference event.