AI-Powered Image Recognition in FileMaker (Part 1 of 2)

Integrating AI-powered image recognition into FileMaker unlocks entirely new ways to interact with your data. Instead of relying solely on manual tagging or rigid folder structures, you can now let the system interpret visual content and surface relevant images through search.

This is Part 1 of a two-part series. Here we’ll cover the foundations: what multi-modal embedding models are, how to configure them in FileMaker 2025, and how to generate your first image embeddings. In Part 2, we’ll build the full semantic search workflow.

What Is AI Image Recognition in FileMaker?

FileMaker 2025 introduces multi-modal embedding models — AI models that can process both text and images, converting them into numerical representations (vectors) that capture their meaning.

This means FileMaker can now:

• Understand image content without manual tagging
• Search for images using text descriptions (e.g., “find photos of damaged roofing”)
• Find visually similar images based on a reference photo
• Bridge text and visual data in a single search operation

This isn’t traditional image recognition (like a binary “is there a face in this photo?”). It’s semantic understanding — the system grasps what an image depicts at a conceptual level.

How Multi-Modal Embeddings Work

The concept is elegantly simple:

1. An image is processed by the AI model
2. The model outputs a vector (a list of numbers) that represents the image’s meaning
3. Text can be processed by the same model into a comparable vector
4. You can compare vectors to find matches between text queries and images

Because both text and images are converted to the same vector space, you can search across modalities. A text query like “red sports car” will match images of red sports cars — even if those images were never tagged or labeled.

Setting Up Your AI Model Account

Before you can generate embeddings, you need to configure an AI model account in FileMaker.

Step 1: Choose Your Provider

FileMaker 2025 supports custom model endpoints. For multi-modal embeddings, you’ll need a provider that offers CLIP-based or similar models. Common options include:

• OpenAI (text embeddings, with image support through the API)
• Custom hosted models (e.g., CLIP via a self-hosted API)
• Cloud AI services (Google, AWS, Azure)

Step 2: Configure the AI Account

Use the Configure AI Account script step:

Configure AI Account [
    Account Name: "image-embedding"
    Model Provider: Custom
    API Key: $apiKey
    Endpoint URL: $endpointUrl
]

Step 3: Set Up the AI Model

Configure the specific model you’ll use for image embeddings:

Configure AI Model [
    Model Name: "clip-embedding"
    Account: "image-embedding"
    Model Type: Multi-Modal Embedding
    Model ID: "clip-vit-base-patch32"
]

The model type must be set to Multi-Modal Embedding to support both text and image inputs.

Generating Your First Image Embedding

With the model configured, generating an embedding is straightforward:

Set Variable [ $embedding ; Value:
    AIModelEmbedding( "clip-embedding" ; Images::photo_container )
]
Set Field [ Images::embedding_vector ; $embedding ]
Commit Records/Requests

The AIModelEmbedding function:

• Takes the model name and a container field as inputs
• Sends the image to the configured model endpoint
• Returns a vector (stored as text in FileMaker)

What Gets Stored

The embedding vector is a long string of numbers — typically 512 or 768 dimensions depending on the model. It looks something like:

[0.0234, -0.1567, 0.0891, 0.2341, ...]

This is what FileMaker uses for semantic comparison. You don’t need to understand the individual numbers — the model handles the meaning.

Preparing Your Data

Container Fields

Your images should be stored directly in container fields (not as external references) for most reliable embedding generation. Supported formats include JPEG, PNG, and TIFF.

Embedding Storage

Create a text field to store the embedding vector. This field should be:

• Indexed for search performance
• Large enough to hold the full vector (set to unlimited storage)
• Not displayed to users (it’s machine-readable, not human-readable)

Metadata Fields

Consider adding fields for:

• Embedding status — Has this image been processed?
• Embedding date — When was the embedding generated?
• Model version — Which model produced this embedding? (Important if you switch models later)

Responsible Use Considerations

Data Privacy

When you send images to an AI model for embedding:

• Where is processing happening? Cloud endpoints mean your images leave your server
• What does the provider do with your data? Read the terms of service carefully
• Are there sensitive images? Medical images, personal photos, or confidential documents require extra care

Model Limitations

Multi-modal models have known limitations:

• Cultural bias — Models trained primarily on Western image datasets may not perform equally well on images from other cultural contexts
• Accuracy varies — Abstract, artistic, or highly specialized images may not embed well
• No perfect matches — Semantic search returns similar results, not exact matches. Always present results as suggestions

Transparency

Document your AI image workflow:

• Which model are you using?
• Where are images processed?
• Who has access to the embedding data?
• How are search results used in decision-making?

What’s Next

In Part 2, we’ll take these embeddings and build a complete semantic image search system — including the search interface, result ranking, and practical optimization tips.

If you’re already working with AI in FileMaker and want guidance on responsible implementation, schedule a free call.