Module classifier 

ICE Classifier / Router

Implements the Interpreter-Classifier-Executor cognitive architecture inspired by OpenAgent’s cortex design.

Overview

Before sending a large prompt to the LLM, the classifier:

1. Interprets the user’s intent from their message
2. Classifies the task type (coding, research, file ops, etc.)
3. Routes only the relevant tools into the LLM context

This reduces token usage and prevents “tool overload” where having too many tools degrades LLM performance.

Architecture

User Input
    │
    ▼
┌──────────────┐
│ Interpreter  │  Extract intent + entities
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Classifier   │  Map intent → task category
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Router       │  Filter tools for category
└──────────────┘

Example

use bob_runtime::classifier::{Classifier, HeuristicClassifier};

let classifier = HeuristicClassifier::new();
let routing = classifier.classify("read the file src/main.rs", &tools).await;
// routing.filtered_tools only contains file-related tools

Structs

ClassificationResult

Classification result with confidence and routing information.

HeuristicClassifier

Heuristic keyword-based classifier.

PassThroughClassifier

No-op classifier that returns all tools without filtering.

Enums

TaskCategory

Task categories for classification.

Traits

Classifier

Trait for task classifiers.