Composite embedders

[dureuill]

Feature name	How to enable	Description	Missing for stabilization	Expected stabilization date/version
composite embedders	curl \ -X PATCH 'http://localhost:7700/experimental-features/' \ -H 'Content-Type: application/json' \ --data-binary '{ "compositeEmbedders": true }'	Enables using a different embedder at search and at indexing time	Confidence in the API surface and the performance of local embedders when many search requests are performed	N/A

Meilisearch v1.14 (to be released on April 14th 2025) introduces composite embedders:

• A new source of embedders called "composite", with two new parameters: searchEmbedder and indexingEmbedder, to specify two embedders to be used respectively at search/indexing time.
• Please refer to public usage for details

How to use the feature?

Please refer to the public API page

What is an experimental feature

By enabling this feature via the /experimental-features route, you opt into the following:

• The API and the behavior of the way composite embedders are declared and work can break between two minor versions of Meilisearch.
• The feature could be removed altogether in a future minor version of Meilisearch

You can use this feature in production, but be prepared to update your code from one version to the next.

Why is this feature not stable yet?

• We are refining the API and making sure it's not too complex or too simple
• We are testing the performance of local embedders in a variety of real-life scenarios

🗣️ You are welcome to give feedback about the score details or ask any question on its usage; we are eager to collect feedback on the feature

⚠️ Disabling the feature

To fully disable the feature, you need to delete all the composite embedders in you embedders settings. To remove an embedder, set it to null in a PATCH settings call: { "embedders: { "YOUR_EMBEDDER_NAME": null } }.

When all composite embedders have been removed from all indexes, call the experimental feature route to set compositeEmbedders to false.

[PierreMesure]

Hi, thank you so much for your work on Meilisearch. The vector/hybrid search has a great developer UX.

I ran into an issue regarding composite embedders as I am currently maintaining a database with a large number of already chunked and embedded documents. I am importing them in a Meilisearch index so my embeddings are userProvided but I was hoping I could use a composite embedder to make Meilisearch use the same API I used for embedding to compute a vector for my query.

Turns out it's not possible to have a composite embedder with userProvided for the indexer. Or I could let Meilisearch regenerate all the embeddings but that's unnecessary when they already exist. So instead, my frontend is calling my embedding API and the query vector is then provided to Meilisearch API. It works fine but it would be great to let Meilisearch handle the embedding. 😊

Let me know if you want a separate discussion thread for this specific issue.

[dureuill]

Hello,

thanks for the kind words; if I understand your use case well, you want to have an embedder that is called by Meilisearch at search time, and provide the embeddings at indexing time in your document.

If that is your use case, you can implement it without a composite embedder.

1. In a fresh index (no documents), add a new embedder with any name to the settings/embedders object. Let's say myEmbedder. Configure myEmbedder so that it can reach your actual embedder provider from its REST API. For the documentTemplate, sets something that will reliably fail to embed, so as to make sure that you cannot accidentally call the embedder at indexing time. For instance: documentTemplate: "{{doc.if_you_see_this_Meilisearch_attempted_to_generate_an_embedding_at_indexing_time}}".
2. Add your documents to the index, passing the key _vectors: { myEmbedder: [...embedding for the document...] } } to each document, in the same way as you would for a userProvided embedder.
3. Meilisearch will not attempt to generate embeddings for your documents, as they are manually provided.
4. At search time, pass your query as normal (without an explicit vector) with hybrid: {embedder: "myEmbedder"}, and Meilisearch will use your embedder to embed the query without you having to do it manually.

Actually, explicitly passing embeddings to documents is always available regardless of the embedder. Under the hood, all that userProvided do is make sure to error if Meilisearch has to embed a document. Since you can emulate that behavior at indexing time with a failing document template, you can use any embedder that will work at search time and be skipped at indexing time as long as you manually provide the embeddings in the documents.

🌱 I wrote this message without using any genAI tools

[PierreMesure]

Fantastic, I'll try soon and let you know how it went!

[PierreMesure]

I have tried and it works great! I think it would be better to clarify the documentation.

[PierreMesure]

Thank you for your answer!