Multi2vec-clip without storing image


I’m currently trying to store images in weaviate using the multi2vec-clip module without actually storing the blob images, I want to keep them in S3 and just use weaviate for indexing and searching.

I’ve successfully achieved this by calling the multi2vec-clip container to manually vectorize each image and store it in weaviate using the withProperties + withVector operators.

However, I had to define at least one textField in my schema, otherwise I wouldn’t be able to use this module. I still want to use the module because I don’t want to manually vectorize my prompts and without this module I can’t use the nearText or nearImage query operators.

I configured the description as a textField which contains a detailed generated description of each image.

My question is, how is this field being used, if it is at all?
When I do a query using nearText, does it vectorize the prompt and compare against the vector or is it using somehow the description field, like a combination of both vector + description?

Is there a better way to achieve this: manually generating the image vector at import time, but use the module vectorizer for the query prompt?

Should I be using a different module instead?

Server Setup Information

  • Weaviate Server Version: 1.14.1
  • Deployment Method: local docker
  • Multi Node? Number of Running Nodes: 1
  • Client Language and Version: Javascript/Typescript

Any additional Information


version: '3.4'
    restart: on-failure:0
      - "8080:8080"
      LOG_LEVEL: "debug"
      DEFAULT_VECTORIZER_MODULE: multi2vec-clip
      CLIP_INFERENCE_API: "http://multi2vec-clip:8080"
      ENABLE_MODULES: "multi2vec-clip"

    image: semitechnologies/multi2vec-clip:sentence-transformers-clip-ViT-B-32-multilingual-v1-1.2.7
      - 8081:8080

collection schema

      "class": "StockImage",
      "moduleConfig": {
          "multi2vec-clip": {
              "textFields": [
      "vectorIndexType": "hnsw",
      "properties": [
          "dataType": [
          "name": "filename"
          "dataType": [
          "name": "url"
          "dataType": [
          "name": "description"

hi @ffleandro ! Welcome to our community :hugs:

There is a feature request that I believe it suits this use case:

Please, consider leaving a thumbs up so we can measure it’s popularity and move it to be a planned feature in our roadmap

Also, there are some internal discussion around a configuration/way that could avoid having the blob. It is needed - at least for now - as if you change a vectorizable text, a new vector will be produced…

As you were able to manually vectorize your images/texts and provide the vectors while inserting the object, Weaviate will not vectorize the image for you.

One thing you are probably missing here, while generating your own vectors: Whenever you have texts and images properties, Weaviate will pass all the vectorizable properties to the inference model (like in here) and once it receives those back, it will combine all those vectors, even taking into account some configurable weights, as you can see with this code here

With that weighted combined vector and having the multi2vec-clip vectorizer configured, whenever you do a nearText or nearImage query, Weaviate will vectorize that prompt and match it against the vectors your have and produce close results to all modalities.

As you probably has only vectorized the image, and not combined those with with your texts vectors, your nearText will probably not return close results to your objects :thinking:

AFAIK, this is the best approach for this case, where you don’t want to store the blob in Weaviate, while we don’t implement the aforementioned feature request.

However, as mentioned, if you provided only an image vector, you will only be able to perform nearImage queries.

Let me know if this helps :slight_smile:


by the way, 1.14.1 is a really old version.

We strongly recommend upgrading to a newer version :slight_smile: