EmbeddingToolSearchStrategy

This docs was updated at: 2026-03-21

com.paragon.agents.toolsearch.EmbeddingToolSearchStrategy  ·  Class

Implements ToolSearchStrategy

---

A tool search strategy that uses embedding vectors for semantic similarity matching.

This strategy embeds both the query and tool metadata (name + description) using an EmbeddingProvider, then ranks tools by cosine similarity to the query embedding. This enables semantic matching — e.g., matching "temperature outside" to a tool named "get_weather" even though they share no keywords.

Note: This strategy makes an API call to the embedding provider on each search, so it has higher latency than BM25ToolSearchStrategy or RegexToolSearchStrategy. For best performance, tool embeddings are computed once and cached.

Example

EmbeddingProvider provider = new OpenRouterEmbeddingProvider(apiKey);
ToolSearchStrategy strategy = new EmbeddingToolSearchStrategy(provider, "text-embedding-3-small", 5);
Agent agent = Agent.builder()
    .name("Assistant")
    .toolRegistry(ToolRegistry.builder()
        .strategy(strategy)
        .deferredTools(hundredsOfTools)
        .build())
    .build();

See Also

• ToolSearchStrategy
• EmbeddingProvider

Since: 1.0

Methods

EmbeddingToolSearchStrategy

public EmbeddingToolSearchStrategy(
      @NonNull EmbeddingProvider embeddingProvider, @NonNull String model, int maxResults)

Creates a new embedding-based tool search strategy.

Parameters

Name	Description
embeddingProvider	the provider for creating embeddings
model	the embedding model identifier (e.g., "text-embedding-3-small")
maxResults	maximum number of tools to return

---

EmbeddingToolSearchStrategy

public EmbeddingToolSearchStrategy(
      @NonNull EmbeddingProvider embeddingProvider, @NonNull String model)

Creates a new embedding-based tool search strategy with a default limit of 5.

Parameters

Name	Description
embeddingProvider	the provider for creating embeddings
model	the embedding model identifier

---

maxResults

public int maxResults()

Returns the maximum number of results this strategy returns.

Returns

the max results limit

---

invalidateCache

public synchronized void invalidateCache()

Invalidates the cached tool embeddings, forcing recomputation on the next search.

Call this if the set of deferred tools changes after construction.