Add the ability to dynamically rerank the search results based on Cohere using a personalized prompt. For more information on how to set it up, see the dedicated experimental feature discussion.

Add the ability to upload snapshots directly to S3. It has many advantages, such as being able to stream the entire process and effectively utilizing multipart technology to send chunks of data in parallel. For more information on how to use it, see the dedicated experimental feature discussion.

The value of the Authorization header is now redacted when getting webhook, getting webhooks, or in the object returned when posting a new webhook or deleting a webhook.

• Impact on Cloud: Similarly to embedder API, the UI should make sure to never send the Authorization header back after it has been redacted
• Technically a breaking change, as users could previously get the key value back when getting the webhook, and that will no longer be possible

This bug was causing crashes in the recent indexer optimizations. If you deactivated these optimizations by setting the following environment variables:

MEILI_EXPERIMENTAL_NO_EDITION_2024_FOR_PREFIX_POST_PROCESSING=true
MEILI_EXPERIMENTAL_NO_EDITION_2024_FOR_FACET_POST_PROCESSING=true

You can now safely reactivate them without experiencing memory leaks.

German word segmentation relies on a word dictionary to segment words, but if a word is not part of the dictionary, it is cut into bigrams. The segmenter will now skip segmenting unknown words:

Source: "Feuchteschutz"
Before: ["Fe" "uc" "ht" "es, "ch", "utz"]
After:  ["Feuchteschutz"]

Source: "insgesamt"
Before: ["in" "sg" "es" "amt"]
After:  ["insgesamt"]

If you have a Meilisearch database containing German words, you must reindex your data manually.

It’s very common for Chinese, numbers, and English to appear together in the same sentence. We now ensure that numbers and English are not segmented differently between segmenters:

Source: "我从2025年开始学习Rust语言。"
Before: ["我", "从", "2", "0", "2", "5", "年", "开始", "学习", "R", "u", "s", "t", "语言", "。"]
After:  ["我", "从", "2025", "年", "开始", "学习", "Rust", "语言", "。"]

If you have a Meilisearch database containing Chinese words, you must reindex your data manually.

• Add Flickr example to README by @Kerollmops in https://github.com/meilisearch/meilisearch/pull/5961

Thanks to @JinheLin, @dureuill, @Kerollmops, and @ManyTheFish for their contribution! 🎉

This release features some improvements with the interaction of the vector store and the searchCutoffMs when using the vectorStore: experimental. It also introduces the metadata header Meili-Include-Metadata on the search request that adds a metadata field to the response. These metadatas contains one uid by query and a reminder of the indexUid and its primary key. We also introduced minor bug fixes around the compaction to improve the interaction with task cancellation.

• Improve the vector store search cutoff by @dureuill in https://github.com/meilisearch/meilisearch/pull/5945
• Improve compaction behaviors by @Kerollmops in https://github.com/meilisearch/meilisearch/pull/5946
• Search metadata by @ManyTheFish in https://github.com/meilisearch/meilisearch/pull/5926

• Adapt the standards of prototypes by @curquiza in https://github.com/meilisearch/meilisearch/pull/5942
• Bump Dockerfile alpine version to 3.22 by @PedroTroller in https://github.com/meilisearch/meilisearch/pull/5866

• @PedroTroller made their first contribution in https://github.com/meilisearch/meilisearch/pull/5866

Full Changelog: https://github.com/meilisearch/meilisearch/compare/v1.23.0...v1.24.0

This release introduces a new compact route on the index routes, which appends a new compaction task to the queue. Meilisearch uses an LMDB environment by index, and indexes start to fragment after some time. We have noticed that the indexes generally have 30% fragmentation. By defragmenting the environment, we've seen large (2-4x) speed-ups in terms of search and indexation. This is primarily due to the reordering of the LMDB internal pages and the removal of scattered free pages throughout the file, thereby relocating the content to the beginning.

We also worked on parallelizing the post-processing of facets. We noticed that a lot of time was spent iterating over the prefixes of the index in a single-threaded loop. We redesigned this part of the indexation to make it multi-threaded. We have seen a 4x and 6x improvement in terms of time spent on this operation.

• Introduce a task to compact an index by @Kerollmops in https://github.com/meilisearch/meilisearch/pull/5929
• Parallelize bulk facets & word prefix fid/position docids by @Kerollmops in https://github.com/meilisearch/meilisearch/pull/5307
• Change Java version in SDK CI by @curquiza in https://github.com/meilisearch/meilisearch/pull/5910
• Minor improvement in OpenAPI CI by @curquiza in https://github.com/meilisearch/meilisearch/pull/5834
• Add request uid to search routes by @ManyTheFish in https://github.com/meilisearch/meilisearch/pull/5863

• Fix ranking score bug when sort is present by @ManyTheFish in https://github.com/meilisearch/meilisearch/pull/5933
• Synonym performance fix by @ManyTheFish in https://github.com/meilisearch/meilisearch/pull/5930
• Update README.md to fix newsletter link by @EazyAl in https://github.com/meilisearch/meilisearch/pull/5911
• Try to fix GH license detection again by @dureuill in https://github.com/meilisearch/meilisearch/pull/5938

• Remove release-drafter and encourage usage of GitHub-generated notes by @curquiza in https://github.com/meilisearch/meilisearch/pull/5935
• Show Dependabot dependency upgrade in the changelog by @curquiza in https://github.com/meilisearch/meilisearch/pull/5900
• Bump actions/setup-go from 5 to 6 by @dependabot[bot] in https://github.com/meilisearch/meilisearch/pull/5912
• Bump actions/setup-dotnet from 4 to 5 by @dependabot[bot] in https://github.com/meilisearch/meilisearch/pull/5914
• Bump actions/setup-node from 4 to 5 by @dependabot[bot] in https://github.com/meilisearch/meilisearch/pull/5915
• Bump sigstore/cosign-installer from 3.9.2 to 3.10.0 by @dependabot[bot] in https://github.com/meilisearch/meilisearch/pull/5916
• Bump actions/setup-python from 5 to 6 by @dependabot[bot] in https://github.com/meilisearch/meilisearch/pull/5913

• @EazyAl made their first contribution in https://github.com/meilisearch/meilisearch/pull/5911

Full Changelog: https://github.com/meilisearch/meilisearch/compare/v1.22.1...v1.23.0

This version contains a minor fix that affects remote federated search users. If you are not a remote federated search user, it is not necessary to migrate from v1.22.x.

• v1.22.2 raised the remote federated search timeout for waiting nodes from 5 to 30s. This version makes it configurable by setting the environment variable MEILI_EXPERIMENTAL_REMOTE_SEARCH_TIMEOUT_SECONDS to a positive integer value. Please note that no CLI flag or configuration entry is available. By @dureuill in https://github.com/meilisearch/meilisearch/pull/5932

• Remote federated search would declare remotes as "timeout'd" after 5s. This timeout was raised to 30s to accomodate some search configurations. By @dureuill in https://github.com/meilisearch/meilisearch/pull/5919

• use the latest version of zerometry that supports collection, lines and multi-lines (#5907) @irevoire
• Delete oldest tasks first (#5906) @dureuill

❤️ Huge thanks to our contributors: @dureuill and @irevoire.

• Introduce a new geo backend to store geojson and filter on polygon
  1. Make the _geojson field filterable
  2. Then send your documents with a _geojson field filled with a valid geojson
  3. Filter your documents with the new _geoPolygon filter, or the old _geoBoudingBox and _geoPoints filter

• Document template: Correctly render when indexing first item in array by @dureuill in https://github.com/meilisearch/meilisearch/pull/5896
• arroy to hannoy conversion fails with binary quantized distances by @nnethercott https://github.com/meilisearch/meilisearch/pull/5891

❤️ Huge thanks to our contributors: @nnethercott, @Kerollmops, @ManyTheFish, @dureuill and @irevoire.

Full Changelog: https://github.com/meilisearch/meilisearch/compare/v1.21.0...v1.22.0

• Introduce a new vector store backend for better performance, especially if using the binary quantization
  1. Enable the new vectorStoreSetting experimental feature
  2. Then change the vectorSetting index setting to "experimental" for the indexes where you want to try the new vector store
  • Done in #5767 by @Kerollmops
• Add Persian support (update charabia to v0.9.7) (#5848) @ManyTheFish

• Observing the progress trace during indexing no longer removes parts of the trace (#5884) @irevoire
• Fix dumpless upgrade decoding error when upgrading with a rest embedder (#5886) @dureuill.
  • In case you had encountered the issue, use the dumpless upgrade to v1.21 to fix it.

❤️ Huge thanks to our contributors: @ja7ad, @agourlay, @Kerollmops, @ManyTheFish, @dureuill and @irevoire.

• Display the progressTrace in in-progress batches (#5858) @shreeup

• Send the version when returning prometheus metrics (#5876) @irevoire

• Bump tracing-subscriber from 0.3.19 to 0.3.20 (#5869) @dependabot[bot]

• Fix scheduled CI failure (#5856) @arithmeticmean

❤️ Huge thanks to our contributors: @ManyTheFish, @arithmeticmean, @curquiza, @dureuill, @irevoire, @shreeup and dependabot[bot].

In previous versions of Meilisearch, mixing hybrid search with filters, as shown below, could multiply the search time by hundreds.

{
  "q": "hello world",
  "limit": 100,
  "filter": "tag=science"
  "hybrid": {
    "semanticRatio": 0.5,
    "embedder": "default"
  }
}

Meilisearch will now directly compute the semantic distance with the filtered candidates if only a few candidates come from the filter, instead of searching for the closest embeddings matching the filter in the vector database.