Embedding model

An EmbeddingModel describes the dimensions, data type, and distance metric of vectors stored in an index. The schema layer uses it as the source of truth to size embedding tensors and select the ranking metric, independently of the embedder that produces vectors at runtime.

The embedder you pass to the Pipeline (see Embedders) produces the vectors; the EmbeddingModel on the schema describes the index. The two must agree on dimensions.

There are two flavors, discriminated by type:

• MistralEmbeddingModel: a model served by the Mistral Embedding API. Use this type to instantiate a MistralEmbedder. Search Toolkit validates dimensions against known presets. A dimension that doesn't match the model name raises DimensionMismatchError. Pass a MistralEmbeddingPreset to create_schema to use its defaults, or build it explicitly to override the data type or distance metric:

  from mistralai.search.toolkit.embedding import (
      DistanceMetric,
      MistralEmbeddingModel,
      MistralEmbeddingPreset,
      VectorDType,
  )
  
  # Pass a preset directly to create_schema (dimensions and metric filled in)
  from mistralai.search.toolkit.plugins.vespa.migration import create_schema
  
  create_schema(
      ...,
      embedding_model=MistralEmbeddingPreset.MISTRAL_EMBED_DIM_1024,
  )
  
  # Build explicitly to override the dtype or distance metric
  model = MistralEmbeddingPreset.MISTRAL_EMBED_DIM_1024.build_embedding_model(
      dtype=VectorDType.FLOAT16,
      distance_metric=DistanceMetric.INNER_PRODUCT,
  )
  
  # Construct directly (dimensions must match the known preset for that name)
  model = MistralEmbeddingModel(name="mistral-embed", dimensions=1024)

  from mistralai.search.toolkit.embedding import (
      DistanceMetric,
      MistralEmbeddingModel,
      MistralEmbeddingPreset,
      VectorDType,
  )
  
  # Pass a preset directly to create_schema (dimensions and metric filled in)
  from mistralai.search.toolkit.plugins.vespa.migration import create_schema
  
  create_schema(
      ...,
      embedding_model=MistralEmbeddingPreset.MISTRAL_EMBED_DIM_1024,
  )
  
  # Build explicitly to override the dtype or distance metric
  model = MistralEmbeddingPreset.MISTRAL_EMBED_DIM_1024.build_embedding_model(
      dtype=VectorDType.FLOAT16,
      distance_metric=DistanceMetric.INNER_PRODUCT,
  )
  
  # Construct directly (dimensions must match the known preset for that name)
  model = MistralEmbeddingModel(name="mistral-embed", dimensions=1024)

• CustomEmbeddingModel: any embedding model not served by the Mistral Embedding API, such as a third-party or self-hosted model. Use this type with your own Embedder implementation. You set the dimensions, data type, and distance metric. Search Toolkit doesn't apply preset validation:

  from mistralai.search.toolkit.embedding import CustomEmbeddingModel, DistanceMetric, VectorDType
  
  model = CustomEmbeddingModel(
      name="my-provider/text-embedding-3-small",
      dimensions=1536,
      dtype=VectorDType.FLOAT32,
      distance_metric=DistanceMetric.COSINE,
  )

  from mistralai.search.toolkit.embedding import CustomEmbeddingModel, DistanceMetric, VectorDType
  
  model = CustomEmbeddingModel(
      name="my-provider/text-embedding-3-small",
      dimensions=1536,
      dtype=VectorDType.FLOAT32,
      distance_metric=DistanceMetric.COSINE,
  )

The discriminated union is exposed as EmbeddingModel, which accepts either flavor.

MistralEmbeddingPreset enumerates the supported Mistral embedding models and carries each one's full name, dimensions, and default distance metric:

Resolve a preset from a model name with MistralEmbeddingPreset.from_name("mistral-embed").

Pair each preset with the matching MistralEmbedder constant on the pipeline:

Every EmbeddingModel carries three vector-config fields, also available as the VectorConfig protocol:

The schema layer uses these to size the embedding tensor and select the ranking metric (for example, cosine similarity vs. inner product).

The EmbeddingModel on the schema describes the index; the embedder you pass to the Pipeline produces the vectors at query and ingestion time. They are separate concerns that must agree on dimensions:

• A MistralEmbeddingModel on the schema pairs with a MistralEmbedder using the matching MODEL_*_EMBEDDING constant.
• A CustomEmbeddingModel on the schema pairs with a custom Embedder.

Keep the dimensions in sync. The schema rejects vectors of the wrong size.

• Embedders: the runtime side that produces vectors.
• Search index: where the schema and the embedding model live.
• Migration helpers: create_schema(embedding_model=...) reference.