Struct ToolSelector 

pub struct ToolSelector { /* private fields */ }

Selects 3–5 relevant tools per task using TF-IDF + cosine similarity, with Jaccard as a fallback for very short descriptions.

Selection rules

• Compute TF-IDF vectors for each tool description at registration time.
• At query time, compute the TF-IDF vector for the intent and score via cosine.
• Sort tools by descending similarity score.
• Return between 3 and min(5, tool_count) tools.
• If no tool has similarity > threshold, return the default tool set instead.

Implementations

impl ToolSelector

pub fn new(_model_path: &Path, preset: &Preset) -> Result<Self>

Create a new ToolSelector from a Preset.

model_path is accepted for API compatibility but is unused.

pub fn register_tools(&mut self, tools: &[ToolDefinition]) -> Result<()>

Register a slice of tools, computing TF-IDF vectors for each description.

pub fn select(&self, intent: &str, max_tools: usize) -> Result<Vec<ToolId>>

Select between 3 and min(5, tool_count) tools whose descriptions best match intent.

Uses TF-IDF + cosine similarity for scoring. Falls back to Jaccard for very short intents (< 3 words) where TF-IDF has insufficient signal.

Returns the default tool set when no tool exceeds the similarity threshold.

pub fn update_tool(&mut self, tool: &ToolDefinition) -> Result<()>

Re-embed a single tool on description change.