Meilisearch

Meilisearch v1.51 is out, and it does two things every search team will appreciate. First, speed. Search requests on documents with many fields are now up to 5.4x faster. One production workload dropped from 153 ms to 28 ms per request - same hardware, same data, no changes needed on the customer's side. Second, smarter merchandising. Our dynamic search rules already let you pin the right products at the top of results. They can now activate based on filters: pin your 👚 "red shirt on sale" product when a shopper filters on color=red and category=shirt, and keep it out of the way everywhere else. You define the facet values once, and the rule fires only when the query's filter resolves to them. Also in this release: dumpless upgrades are now stable, so you can move to a new version without exporting your data. Check the release notes for more instructions on rollout.

Meilisearch or Manticore - which search engine is the better fit for you? In our latest guide, you’ll learn: 💡 How Meilisearch and Manticore differ in architecture and performance 💡 The pros and limitations of each platform 💡 Which tool is better for startups, SaaS apps, and large datasets 💡 Key factors to consider before choosing a search engine Link to full blog below ->

Meilisearch now has a certified plugin for Kestra, created and maintained by the core team. Keeping a search index in sync usually falls to a cron script that someone has to babysit. When it breaks quietly, results go stale until a user notices. With the certified plugin, indexing into Meilisearch runs as a task inside your Kestra workflows – logged, retried on failure, scheduled, and observable. A failed sync becomes something you can see and fix, not a silent gap in your results. No glue code to maintain, just extract → convert → index. We built step-by-step guides for common sources: PostgreSQL, MongoDB, Kafka, Shopify, and more, so you start from a ready-made Kestra workflow.

An AI hallucination in a chatbot is annoying. In a clinical setting, it can harm patients. That's why healthcare teams are turning to RAG (retrieval-augmented generation). Instead of letting the model guess, RAG pulls answers from trusted sources – clinical guidelines, PubMed research, electronic health records – before generating a response. The result: fewer errors, more accurate answers, and AI that clinicians can actually trust. Our latest guide covers how medical RAG works, real use cases (clinical decision support, research search, drug lookups), and how to build one step by step. Link in the first comment.

If you've added a lot of synonyms to an index, you may have seen search slow down as the list grew. Our CTO Clément Renault just fixed that in Meilisearch v1.49.0. Synonyms now load only when a word actually matches one, instead of being carried through every query. That means up to 13x faster search, and even bigger gains if you're running large synonym lists (10k+) 📈 No workarounds, no moving synonyms into separate documents. Just faster search out of the box. Link in the comments.

Validating a document template used to be tedious. In Meilisearch v1.48.0, we've added an experimental renderRoute for exactly that. The new POST /render-template route renders any template or fragment on any input and returns the rendered result – so you can test your document templates and fragments before and after configuring an embedder. It covers indexing and search fragments too, including multimodal inputs. And if you send no input, it returns the template or fragment straight from your index settings. The route is gated behind the renderRoute experimental feature, so you enable it before use (link in comments).

Passing your entire knowledge base to an LLM burns tokens, runs up costs, and slows everything down – often without making the answers any better. Context distillation is the fix. You select and filter your data so the model only sees what's relevant to the query. The payoff is concrete: fewer tokens = lower bills, faster responses, and more consistent answers. In our latest piece, we walk through how it works, the common techniques, where it breaks down, and how it differs from model distillation and fine-tuning:

Last week was VivaTech, and it was so good to meet our users, prospects and customers in person. Grabbing a coffee, sketching something on a napkin to explain what you mean - you can't really replace this stuff over Zoom. It's the best way to actually talk with the community, and there's always a bunch of people who are genuinely happy to meet us IRL (Same in the other direction, trust me.) I'm still thinking about the conversations. I'm honestly impressed by how far people have come on agentic workflows. A year ago most of it was demos and slideware. Now they're shipping real things, in production, with real constraints. The pace is wild. And what surprised me is that RAG is still an unsolved problem for a lot of teams. Impossible to trust. Hard to scale. The gap between "it works in a notebook" and "I can put this in front of customers" is still huge for many of them. There's a lot left to build here, and that's exciting. VivaTech was also a great moment to stay close to our partners. Big thanks to the Amazon Web Services (AWS) crew for the conversations and the energy: Rémi FRANCOIS, Nicolas de Place, Gautier Colomb, Julien Dhainaut. Always a pleasure. See you next year. 🚀

VivaTech 2026 is a wrap! Thank you for keeping us busy at the booth! We got to talk search, RAG, and what people are actually trying to ship right now - so exciting to be able to discuss with teams from Bouygues Telecom, Leboncoin.fr, Michelin, Club Med, Ouest-France, LumApps, and plenty more. Huge thanks to the Amazon Web Services (AWS) crew for the energy as well - always good to build alongside you. Some patterns from these discussions: agentic workflows are getting real, with teams moving past the demo phase and putting things into production. And RAG is still the hard part for most of them: getting retrieval you can trust was a huge point of discussion. This is exactly where we love to dig in. See you next year. 🚀 

Day one at VivaTech and the team is officially in the building 🚀 Find Quentin de Quelen, Chiara Souchal 🐜, Matthew Schimke, Boris Piquet & 🐢 Thomas Payet right now at Hall 7.1, Stand 1H42-003! We're here to talk AI-retrieval, semantic search, and what fast, developer-first search actually looks like when you put it to work. Come say hi, see a demo, or just grab a coffee with us. We're around all week!