Skip to content

ToolSearchStrategy

This docs was updated at: 2026-03-21

com.paragon.agents.toolsearch.ToolSearchStrategy  ·  Interface

Strategy for dynamically selecting which tools to include in an LLM API call.

When an agent has many tools, sending all of them in every API call wastes context window tokens and can degrade tool selection accuracy. A ToolSearchStrategy solves this by searching the available tools based on the user's input and returning only the most relevant subset.

This is the framework-level equivalent of Anthropic's server-side tool search feature, but works with any LLM provider (OpenAI, OpenRouter, local models) because the selection happens client-side before the API call.

Built-in Strategies

  • RegexToolSearchStrategy — pattern-based matching on tool names/descriptions
  • BM25ToolSearchStrategy — TF-IDF scoring for relevance ranking
  • EmbeddingToolSearchStrategy — semantic similarity via embedding vectors

Custom Strategies

ToolSearchStrategy custom = (query, tools) -> tools.stream()
    .filter(t -> mySimilarityModel.score(query, t.getName()) > 0.7)
    .toList();

See Also

  • ToolRegistry

Since: 1.0

Methods

List<FunctionTool<?>> search(@NonNull String query, @NonNull List<FunctionTool<?>> allTools)

Searches the available tools and returns the subset most relevant to the query.

Parameters

Name Description
query the user's input text used to determine tool relevance
allTools all deferred tools available for selection

Returns

the subset of tools to include in the API call (may be empty)