Multiple embedders

Meilisearch supports configuring multiple embedders on the same index. Each embedder generates its own set of vectors, and you can target a specific embedder at search time. This lets you combine different search strategies (text, image, semantic) with specialized models for each.

Why use multiple embedders

A single embedder is a good fit when all your searches are the same type. But real applications often need different search modes:

Text + image search: use a text-optimized embedder alongside a multimodal embedder, so users can search with keywords or with images
Precision vs speed: use a large, high-quality model for precise searches and a smaller, faster model for search-as-you-type suggestions
Different quality levels: use a small model at full precision for quick queries and a large model with binary quantization for deep searches
Multilingual: use a language-specific model for your primary language and a multilingual model as a fallback
Federated search: combine full-text, semantic, and image results in a single federated search request, each powered by the best model for its task

Configure multiple embedders

Add multiple keys to the embedders setting. Each key is a named embedder with its own configuration:

curl \
  -X PATCH 'MEILISEARCH_URL/indexes/products/settings/embedders' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer MEILISEARCH_KEY' \
  --data-binary '{
    "text": {
      "source": "openAi",
      "apiKey": "OPEN_AI_API_KEY",
      "model": "text-embedding-3-small"
    },
    "image": {
      "source": "rest",
      "url": "https://api.voyageai.com/v1/multimodalembeddings",
      "apiKey": "VOYAGE_API_KEY",
      "indexingFragments": {
        "poster": {
          "value": {
            "content": [
              { "type": "image_url", "image_url": "{{doc.poster_url}}" }
            ]
          }
        }
      },
      "searchFragments": {
        "image": {
          "value": {
            "content": [
              { "type": "image_url", "image_url": "{{media.image}}" }
            ]
          }
        }
      },
      "request": {
        "inputs": ["{{fragment}}", "{{..}}"],
        "model": "voyage-multimodal-3"
      },
      "response": {
        "data": [{ "embedding": "{{embedding}}" }, "{{..}}"]
      }
    }
  }'

This configures two embedders: text for keyword-aware semantic search and image for visual similarity search.

Search with a specific embedder

Specify which embedder to use with the hybrid.embedder parameter:

# Semantic text search
curl -X POST 'MEILISEARCH_URL/indexes/products/search' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer MEILISEARCH_KEY' \
  --data-binary '{
    "q": "comfortable running shoes",
    "hybrid": {
      "embedder": "text",
      "semanticRatio": 0.5
    }
  }'

# Image search
curl -X POST 'MEILISEARCH_URL/indexes/products/search' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer MEILISEARCH_KEY' \
  --data-binary '{
    "media": {
      "image": "https://example.com/shoe.jpg"
    },
    "hybrid": {
      "embedder": "image",
      "semanticRatio": 1.0
    }
  }'

Combine embedders with federated search

The most powerful use case for multiple embedders is federated search. You can run full-text, semantic, and image searches in a single request and merge the results:

curl -X POST 'MEILISEARCH_URL/multi-search' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer MEILISEARCH_KEY' \
  --data-binary '{
    "federation": {},
    "queries": [
      {
        "indexUid": "products",
        "q": "running shoes",
        "hybrid": {
          "embedder": "text",
          "semanticRatio": 0.0
        },
        "federationOptions": { "weight": 1.0 }
      },
      {
        "indexUid": "products",
        "q": "running shoes",
        "hybrid": {
          "embedder": "text",
          "semanticRatio": 1.0
        },
        "federationOptions": { "weight": 0.8 }
      },
      {
        "indexUid": "products",
        "media": {
          "image": "https://example.com/shoe.jpg"
        },
        "hybrid": {
          "embedder": "image",
          "semanticRatio": 1.0
        },
        "federationOptions": { "weight": 0.5 }
      }
    ]
  }'

This single request combines:

Full-text search (semanticRatio: 0.0) with the highest weight for keyword-relevant results
Semantic text search (semanticRatio: 1.0) for meaning-based matches
Image search using a completely different model for visual similarity

Meilisearch merges all results into one ranked list using the federation weights.

Considerations

Each embedder generates and stores its own vectors. More embedders means more disk usage and longer indexing times.
You can use binary quantization on individual embedders to reduce storage (e.g., quantize the large model but keep the small one at full precision).
Composite embedders can be combined with multiple embedders: use a fast local model for search and a cloud API for indexing, independently for each named embedder.

Next steps

Federated search

Merge results from multiple queries into one ranked list

Binary quantization

Reduce storage for high-dimensional embedders

Image search

Set up multimodal embedders for image search

Choose an embedder

Compare embedding providers for your use case

Capabilities

Full-text search

Hybrid and semantic search

Geo search

Conversational search

Multi-search

Filtering, sorting, and faceting

Personalization

Analytics

Security and tenant tokens

Search rules

Platform

Indexing

Why use multiple embedders

Configure multiple embedders

Search with a specific embedder

Combine embedders with federated search

Considerations

Next steps

Federated search

Binary quantization

Image search

Choose an embedder

​Why use multiple embedders

​Configure multiple embedders

​Search with a specific embedder

​Combine embedders with federated search

​Considerations

​Next steps

Federated search

Binary quantization

Image search

Choose an embedder

Why use multiple embedders

Configure multiple embedders

Search with a specific embedder

Combine embedders with federated search

Considerations

Next steps