Normalizing a vector when using weaviate with a text2vec-transformers

Support
1

I am using weaviate 1.3.1 local via docker with a text2vec-transformers (multilingual-e5-large-inference) image as the backend embedder (see below for details)

From this embedder I read:

* The E5 models (including `e5-multilingual-large`) produce embeddings that are **meant to be L2-normalized** (i.e., turned into unit vectors).
* The **contrastive loss** used during training assumes cosine similarity — which mathematically requires **unit-length vectors** for correct behavior.

and also asking ChatGPT about what happens if you skip it (which I have done until now) it says:

### Consequences of Skipping Normalization:

* Storing **raw, unnormalized vectors** in Weaviate (or any vector DB) means you're not really using **cosine similarity**, even if the DB claims to.
* This can lead to **suboptimal retrieval results**, particularly poor semantic matching and lower recall/precision.

The embedder /vectors endpoint does not allow me to do anything other that sending it a string:

image
image3132×850 139 KB

so the question is “does this container wrongly return the raw E5 vector or does it normalize it under the covers?”

Thank you very much.

Docker compose declaration follows.

I am using weaviate 1.3.1 local via docker as per the following config:

services:
  weaviate131:
    networks:
      - mema_docker_compose_weaviate_net 
    command:
      - --host
      - 0.0.0.0
      - --port
      - '8080'
      - --scheme
      - http
    image: cr.weaviate.io/semitechnologies/weaviate:1.31.0
    ports:
      - "${WEAVIATE_HTTP_PORT:-8099}:8080"
      - "${WEAVIATE_GRPC_PORT:-50099}:50051"
    volumes:
      - weaviate_131_data:/var/lib/weaviate
    restart: on-failure:0
    environment:
      QUERY_DEFAULTS_LIMIT: 25
      AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED: 'true'
      PERSISTENCE_DATA_PATH: '/var/lib/weaviate'
      DEFAULT_VECTORIZER_MODULE: 'text2vec-transformers'
      ENABLE_MODULES: 'text2vec-transformers'
      TRANSFORMERS_INFERENCE_API: 'http://multilingual-e5-transformers:8080'
      CLUSTER_HOSTNAME: 'node1'
    depends_on:
      - multilingual-e5-transformers

  multilingual-e5-transformers:
    build:
      context: .
      dockerfile: multilingual-e5-large.Dockerfile
    image: multilingual-e5-large-inference
    networks:
      - mema_docker_compose_weaviate_net
    ports:
      - "${TRANSFORMER_PORT:-8092}:8080"
    environment:
      ENABLE_CUDA: '0'
    restart: on-failure:0

volumes:
  weaviate_131_data:

networks:
  mema_docker_compose_weaviate_net:
    external: true
2

Ciao amico @rjalex !!! Long time no see!

This is a hard one that I do not have the answer :slight_smile:

I will raise this to our team.

Just to confirm, you referring to this endpoint here, right?

Thanks!

1 Like
3

Hi Duda,
I was almost going to call you Saint Duda since you always help me a lot but then remembered that to be a Saint the prerequisite is that you should be dead :slight_smile: :slight_smile: :slight_smile:

Yes I am using the /vectors endpoint.

What I really hope is that the container dedicated to e5-multilingual does the required normalisation.

Will keep tuned to have a confirmation.

Muito obrigado
Bob