Skip to main content

Optimize tool discovery

When Virtual MCP Server (vMCP) aggregates many backend MCP servers, the total number of tools exposed to clients can grow quickly. The optimizer addresses this by filtering tools per request, reducing token usage and improving tool selection accuracy.

For the desktop/CLI approach using the MCP Optimizer container, see the MCP Optimizer tutorial. This guide covers the Kubernetes operator approach using VirtualMCPServer and EmbeddingServer CRDs.

Benefits​

  • Reduced token usage: Only relevant tools are included in context, not the entire toolset
  • Improved tool selection: The right tools surface for each query. With fewer tools to reason over, agents are more likely to choose correctly

How it works​

  1. You send a prompt that requires tool assistance
  2. The AI calls find_tool with keywords extracted from the prompt
  3. vMCP performs hybrid semantic and keyword search across all backend tools
  4. Only the most relevant tools (up to 8 by default) are returned
  5. The AI calls call_tool to execute the selected tool, and vMCP routes the request to the appropriate backend
How search works internally

The optimizer uses an internal SQLite database for both keyword search (using full-text search) and storing semantic vectors. Keyword search runs locally against this database; semantic search uses vectors generated by an embedding server. You can control how results from these two sources are blended — see the parameter reference for details.

Quick start​

Step 1: Create an EmbeddingServer​

Create an EmbeddingServer with default settings. This deploys a text embeddings inference (TEI) server using the BAAI/bge-small-en-v1.5 model:

embedding-server.yaml
apiVersion: toolhive.stacklok.dev/v1alpha1
kind: EmbeddingServer
metadata:
  name: my-embedding
  namespace: toolhive-system
spec: {}
tip

Wait for the EmbeddingServer to reach the Running phase before proceeding. The first startup may take a few minutes while the model downloads.

kubectl get embeddingserver my-embedding -n toolhive-system -w

Step 2: Add the embedding reference to VirtualMCPServer​

Update your existing VirtualMCPServer to include embeddingServerRef. This is the only change needed to enable the optimizer. When you set embeddingServerRef, the operator automatically enables the optimizer with sensible defaults. You only need to add an explicit optimizer block if you want to tune the parameters.

VirtualMCPServer resource
apiVersion: toolhive.stacklok.dev/v1alpha1
kind: VirtualMCPServer
metadata:
  name: my-vmcp
  namespace: toolhive-system
spec:
  embeddingServerRef:
    name: my-embedding
  config:
    groupRef: my-group
  incomingAuth:
    type: anonymous

Step 3: Verify​

Check that the VirtualMCPServer is ready:

kubectl get virtualmcpserver my-vmcp -n toolhive-system

Look for READY: True in the output. Once ready, clients connecting to the vMCP endpoint see only find_tool and call_tool instead of the full backend toolset.

EmbeddingServer resource​

The EmbeddingServer CRD manages the lifecycle of a TEI server. An empty spec: {} uses all defaults. The two most important fields you can customize are:

  • model: The Hugging Face embedding model to use. The default (BAAI/bge-small-en-v1.5) is the tested and recommended model. You can substitute any embedding model available on Hugging Face — see the MTEB leaderboard to compare options.
  • image: The container image for text-embeddings-inference (TEI). The default is the CPU-only image (ghcr.io/huggingface/text-embeddings-inference:cpu-latest). Swap this for a CUDA-enabled image if you have GPU nodes available.

For the complete field reference, see the EmbeddingServer CRD specification.

ARM64 compatibility

The default text-embeddings-inference (TEI) image (ghcr.io/huggingface/text-embeddings-inference) is amd64-only. If you are running on ARM64, you must pre-pull the image for linux/amd64 and load it into your cluster:

docker pull --platform linux/amd64 \
  ghcr.io/huggingface/text-embeddings-inference:cpu-1.7
kind load docker-image \
  ghcr.io/huggingface/text-embeddings-inference:cpu-1.7

Tune the optimizer​

To customize optimizer behavior, add the optimizer block under spec.config in your VirtualMCPServer resource:

VirtualMCPServer resource
spec:
  config:
    groupRef: my-group
    optimizer:
      embeddingServiceTimeout: 30s
      maxToolsToReturn: 8
      hybridSearchSemanticRatio: '0.5'
      semanticDistanceThreshold: '1.0'

Parameter reference​

ParameterDescriptionDefault
embeddingServiceTimeoutHTTP request timeout for calls to the embedding service30s
maxToolsToReturnMaximum number of tools returned per search (1-50)8
hybridSearchSemanticRatioBalance between semantic and keyword search. 0.0 = all keyword, 1.0 = all semantic. Default gives equal weight to both."0.5"
semanticDistanceThresholdMaximum distance from the search term for semantic results. 0 = identical, 2 = completely unrelated. Results beyond this threshold are filtered out."1.0"
note

hybridSearchSemanticRatio and semanticDistanceThreshold are string-encoded floats (for example, "0.5" not 0.5). This is a Kubernetes CRD limitation, as CRDs do not support float types portably.

EmbeddingServer is always required

Even if you set hybridSearchSemanticRatio to "0.0" (all keyword search), the optimizer still requires a configured EmbeddingServer. The EmbeddingServer won't be used at runtime when the semantic ratio is 0.0, but the configuration must be present due to how the optimizer is wired internally.

Tuning guidance

The defaults are well-tested and work for most use cases. If you do need to adjust them:

  • Lower semanticDistanceThreshold (for example, "0.6") for higher precision: only very close matches are returned
  • Raise semanticDistanceThreshold (for example, "1.4") for higher recall: broader matches are included
  • Increase maxToolsToReturn if the AI frequently cannot find the right tool; decrease it to save tokens
  • Adjust hybridSearchSemanticRatio toward "1.0" if tool names are not descriptive, or toward "0.0" if exact keyword matching is more useful
  • semanticDistanceThreshold filtering is applied before the maxToolsToReturn cap. A low threshold can filter out candidates before the cap takes effect, so you may need to raise the threshold if too few results are returned

Complete example​

This example shows a full configuration with all available options, including high availability for the embedding server, persistent model caching, and tuned optimizer parameters.

The EmbeddingServer runs two replicas with resource limits and a persistent volume for model caching, so restarts don't re-download the model:

embedding-server-full.yaml
apiVersion: toolhive.stacklok.dev/v1alpha1
kind: EmbeddingServer
metadata:
  name: full-embedding
  namespace: toolhive-system
spec:
  replicas: 2
  resources:
    requests:
      cpu: '500m'
      memory: '512Mi'
    limits:
      cpu: '2'
      memory: '1Gi'
  modelCache:
    enabled: true
    storageSize: 5Gi

The VirtualMCPServer uses a shorter embedding timeout (15s) because the EmbeddingServer is co-located with low-latency access. Increase this value if the embedding service is remote or under high load:

vmcp-with-optimizer.yaml
apiVersion: toolhive.stacklok.dev/v1alpha1
kind: VirtualMCPServer
metadata:
  name: full-vmcp
  namespace: toolhive-system
spec:
  embeddingServerRef:
    name: full-embedding
  config:
    groupRef: my-tools
    optimizer:
      embeddingServiceTimeout: 15s
      maxToolsToReturn: 10
      hybridSearchSemanticRatio: '0.6'
      semanticDistanceThreshold: '0.8'
  incomingAuth:
    type: oidc
    oidcConfig:
      type: inline
      inline:
        issuer: https://auth.example.com
        audience: vmcp-example