Meilisearch AI - Boost relevance and UX with fast hybrid and semantic search

@strijdhagen Yes, it's in our plans! We are just about to release the Composite Embedder, which will be a good first step on this path. What are the pain points you want to solve in your case? Mostly, we see pricing or performance pain points that will be addressed, but anything different would be interesting to dive deeper into.

@quentin_dq Mostly lazyness, having everything in one please is awesome :)

@carolina_ferreira2  @maya_shin2  @gmourier  @kui_jason You will love what is coming next; we will be able to run on the same index, the same model on a local CPU for search and on a GPU for indexing. Best of both worlds, it will make Meilisearch the fastest and most scalable hybrid search 🔥

@strift The full team is so excited by this launch 🚀

@dobroslav_dev Thanks for the support! Let's try it; it takes only a few hours to have it running :)

@kay_arkain That means a lot, thank you! I am the product designer responsible for the UI/UX, so it’s really great to hear you like it. 😊 We’re always looking to improve, especially when it comes to AI-specific flows, so if you (or anyone reading this) have any thoughts, we’d love to hear them, via Discord or directly the feedback button in the app.

@faizanjan_ Thanks for the kind words! Great question.

When you perform a hybrid search in Meilisearch, we run two parallel searches: one using our full-text search engine, and the other using vector-based semantic search.

• The full-text search uses Meilisearch’s custom ranking rules to assign a highly precise score to each document based on how well it matches the query terms — better than standard BM25.

• The semantic search uses a KNN (nearest neighbor) search on embeddings to score documents based on how close their meaning is to the query.

What makes Meilisearch stand out is how we combine these two scores into a single, unified ranking. Instead of relying on simple techniques like Fusion Ranking, we apply a smart blending of both relevance signals to ensure the results are not only lexically accurate but also semantically meaningful.

This leads to much more relevant results than using just BM25, just vectors, or even most other hybrid approaches.

@kerollmops might have even more technical detail to add, but that’s the high-level view!

@siahouchangnia Thanks 🤩

Thanks @jonurbonas!

Yes — the ranking rules in Meilisearch are very flexible. You can reorder them, remove the ones you don’t need, or even define custom rules based on your use case — all directly from the Cloud UI.

We also make it easy to experiment: you can test new configurations on a separate index and only apply them to production once you’re confident. Everything is non-destructive, so it’s easy to roll back if needed.

For hybrid search, it’s super simple: you just adjust a ratio to control the balance between semantic and full-text scoring. That gives you fine-grained control over how AI influences the results.