Support for Multiple Vectors per Document in Embeddings

[webzit]

In an experimental setup, I created a Meilisearch instance containing Supreme Court decisions from an EU country. These decisions consistently consist of three main sections:

• A header
• The ruling
• The legal rationale

I've stored these three attributes (among other attributes) in the JSON documents, with individual paragraphs from each section represented as nested objects under these attributes.

I would like to generate and store separate embeddings for each of these paragraphs in Meilisearch. My understanding is that only one embedding per embedding provider can be stored per document. Is there a way to support multiple embeddings per document, or could this functionality be added?

This would be particularly valuable for legal documents where different sections have distinct semantic meanings, enabling more precise semantic search across document components.

(Creating separate documents for all these paragraphs is not an option, as a keyword search would not find the relevant decision if the keywords are spread over several paragraphs of the decision).

[dureuill]

Hello, while Meilisearch does not currently provide a way to do exactly as you request, there are multiple ways you could achieve a close solution

Using multiple embedders

You can define multiple embedders with the same provider and only a different documentTemplate in a single index.
The only drawback is that at search time you'd have to specify which embedder you want to look into. However, using the federated search it should be easy to mix the results from all the embedders

Using a userProvided embedder

A userProvided embedder requires that you send your embeddings manually alongside your documents. In that configuration, you're allowed to send up to 256 embeddings per document and embedder. The drawback to this option is that you would no longer benefit from the integration of Meilisearch to REST embedders, as you would need to generate your embeddings yourself before attaching them to your documents.

[webzit]

Thank you for the very helpful response! I expect this feature will be in demand as embedding use cases grow.

The proposed workarounds are viable but not ideal. Regarding the first approach (multiple embedders + federated search):

• If a document has e.g. 40 embedders with the same provider, would Meilisearch send 40 identical queries to OpenAI (for the same search input), redundantly generating the same embedding 40 times? (I haven't checked the source yet, only if you know it ad hoc)

This inefficiency would make the userProvided embedder more appealing—ideally paired with a custom API wrapper to precompute and inject the embeddings.

[dureuill]

If a document has e.g. 40 embedders with the same provider, would Meilisearch send 40 identical queries to OpenAI (for the same search input), redundantly generating the same embedding 40 times? (I haven't checked the source yet, only if you know it ad hoc)

Meilisearch v1.14, released today, adds a query embedding cache as an experimental feature: https://github.com/orgs/meilisearch/discussions/818

So if you start Meilisearch with --experimental-embedding-cache-entries=150, then assuming all your OpenAI embedders use the same model and dimensions, and all your text queries are identical, then only one call to OpenAI will be made. Furthermore the text queries would also be cached across requests.

This feature is available by default on the Cloud.

[webzit]

Wonderful, thank you!