Microsoft Foundry Embed MCP connector for Power Platform

April 03, 2026

Embeddings turn text and images into numbers that capture meaning. Two sentences about the same topic produce vectors that are close together; unrelated content produces vectors that are far apart. This is the foundation of semantic search, duplicate detection, recommendation engines, and RAG pipelines.

This connector brings Cohere Embed v4 into Power Platform with three tiers of functionality:

• Basic — Compute similarity between any two inputs (text or image) with a single call. No vector store needed.
• Intermediate — Generate raw embedding vectors for text or images. Store and compare them however you want.
• Advanced — Index documents to Azure AI Search with auto-generated vectors and run vector search. A complete RAG retrieval pipeline from Power Automate.

Five MCP tools for Copilot Studio. Five REST operations for Power Automate and Power Apps.

Full source: GitHub repository

Model details

The cross-modal capability is key—text embeddings and image embeddings land in the same vector space. You can compare a product photo against a text description and get a meaningful similarity score.

Tools

MCP tools for Copilot Studio

How similarity works

User: "Are these two product descriptions duplicates?"

1. Agent calls compute_similarity({
     input_a: "Wireless noise-canceling headphones with 30-hour battery",
     input_a_type: "text",
     input_b: "Bluetooth headphones, active noise cancellation, 30h playtime",
     input_b_type: "text"
   })

   → similarity: 0.89
     interpretation: "Very similar"

The connector embeds both inputs, computes cosine similarity, and returns a score with a human-readable interpretation:

Cross-modal comparison

Compare text against images:

User: "Does this product photo match our catalog description?"

Agent calls compute_similarity({
  input_a: "https://storage.blob.core.windows.net/products/headphones.png",
  input_a_type: "image",
  input_b: "Over-ear wireless headphones in matte black with cushioned ear cups",
  input_b_type: "text"
})

→ similarity: 0.74
  interpretation: "Similar"

REST operations for Power Automate and Power Apps

Embed Text

Generate embedding vectors for one or more text strings.

The input_type parameter matters. Use document when embedding content for storage/indexing. Use query when embedding a search query. The model applies different transformations for each to optimize retrieval accuracy.

Embed Image

Generate an embedding vector for an image, optionally paired with text.

Pair an image with text to create a combined embedding that captures both visual and textual meaning.

Compute Similarity

Compare any two inputs without a vector store.

Supports four combinations: text-to-text, text-to-image, image-to-text, image-to-image.

Index Document to AI Search

Embed a document and push it with its vector to Azure AI Search in one call.

Search Similar Documents

Embed a query and run vector search against an Azure AI Search index.

Azure AI Search index requirements

For the IndexDocument and SearchSimilar operations, your AI Search index must have:

• A string field named id (set as the key field)
• A string field named content
• A Collection(Edm.Single) field for vectors (default name: contentVector) with dimensions matching the connector (1024 by default)
• An optional string field named title

Example workflows

Duplicate detection (no AI Search required)

1. User submits two product descriptions
2. ComputeSimilarity returns 0.87 (“Very similar”)
3. Flow flags them as potential duplicates for review

No vector store, no index, no infrastructure—just two strings in, one score out.

RAG pipeline with Azure AI Search

Indexing phase (run once per document):

1. Retrieve document text from SharePoint or Dataverse
2. Call IndexDocument with the text, a unique ID, and your AI Search credentials
3. The connector embeds the text and pushes it with its vector to the index

Query phase (run per user question):

1. User asks a question
2. Call SearchSimilar with the question and AI Search credentials
3. Get back the top 5 most semantically similar documents
4. Send those documents as context to an LLM for answer generation

Cross-modal product matching

1. User uploads a photo of a product
2. ComputeSimilarity compares the image embedding against each product description text
3. Returns the closest matching product by semantic similarity

Prerequisites

1. An Azure subscription with access to Microsoft Foundry
2. Deploy embed-v-4-0 from the Foundry Model Catalog
3. Note the Resource Name and API Key from the deployment
4. (Advanced operations only) An Azure AI Search service with a vector-enabled index

Setting up the connector

1. Deploy Cohere Embed v4

1. Go to the Foundry Model Catalog
2. Select embed-v-4-0 and deploy to your Azure AI Services resource
3. Copy the Resource Name and API Key

2. Create the custom connector

1. Go to Power Platform Maker Portal
2. Navigate to Custom connectors > + New custom connector > Import an OpenAPI file
3. Upload apiDefinition.swagger.json
4. On the Security tab:
   • Authentication type: API Key
   • Parameter label: API Key
   • Parameter name: api-key
   • Parameter location: Header
5. On the Code tab:
   • Enable Code
   • Upload script.csx
6. Select Create connector

3. Test the connector

Test ComputeSimilarity first—it’s the simplest operation and doesn’t need AI Search:

{
  "input_a": "The quick brown fox jumps over the lazy dog",
  "input_b": "A fast auburn fox leaps above a sleeping hound"
}

Verify you get a high similarity score (should be around 0.8+).

4. Add to Copilot Studio

1. In Copilot Studio, open your agent
2. Add this connector as an action—Copilot Studio detects the MCP endpoint via x-ms-agentic-protocol
3. Test with prompts like “How similar are these two descriptions?” or “Search my knowledge base for documents about return policies”

Known limitations

• Text input is limited to 512 tokens per string
• Image embeddings require PNG format; other formats may not work
• One image per embedding call (batch image embeddings not supported)
• AI Search operations pass credentials as parameters (not stored in the connection) because they connect to a separate service
• Cosine similarity scores are relative—they compare two specific inputs, not absolute quality measures
• The default vector dimensions (1024) must match your AI Search index configuration

Files

Resources

• Embed connector source code
• Cohere Embed v4 in Foundry Catalog
• Azure AI Search documentation
• Microsoft Foundry API
• Cohere Embed documentation