Tachyon vs. Meilisearch
How Tachyon and Meilisearch compare on relevance model, storage engine, license, and feature surface.
Tachyon and Meilisearch share a language (Rust) and a target use case (embedded application search), but take different positions on relevance: Meilisearch ships a proprietary, rule-based ranking pipeline tuned for instant-search UX; Tachyon uses BM25, a well-understood, transparent lexical ranking function.
At a glance
| Tachyon | Meilisearch | |
|---|---|---|
| Language | Rust | Rust |
| License | Apache-2.0 | MIT |
| Distribution | Single binary / Docker image | Single binary / Docker image |
| Storage engine | Custom mmap'd segments + WAL | LMDB |
| Relevance | BM25 | Proprietary ranking rule pipeline |
| Typo tolerance | Damerau-Levenshtein, built in | Built in |
| Faceting | Yes | Yes |
| Filtering | Yes, boolean expression language | Yes |
| Sorting | Multi-clause, numeric + relevance | Multi-clause, incl. custom ranking rules |
| Vector / hybrid search | No — see Roadmap | Yes |
| Geo search | No — see Roadmap | Yes |
| Clustering / replication | No — see Roadmap | Enterprise-only |
| Managed cloud offering | No | Meilisearch Cloud |
| Client SDKs | TypeScript, Python, C# | Many, incl. official + community |
Relevance model
Meilisearch ranks with an ordered pipeline of rules (typo count, word proximity, field attribute order, exactness, then your own custom ranking rules) rather than a single scoring function — tuned specifically to make instant-search-as-you-type feel right by default, with limited need to hand- tune. Tachyon uses BM25: one well-documented formula, the same one behind Elasticsearch and Lucene, with per-field weighting on top. See Relevance & BM25 for exactly how it's computed.
Neither is strictly "better" — Meilisearch's pipeline is opinionated and gets instant-search UX right with less configuration; BM25 is more predictable to reason about and more familiar if you're coming from Elasticsearch or Lucene-based systems.
Storage engine
Meilisearch is built on LMDB, a mature memory-mapped B-tree key-value store. Tachyon uses a purpose-built storage layer: a write-ahead log for durability plus immutable, memory-mapped segments merged in the background. Both are mmap-based and avoid holding the full index resident in the process's own heap; the difference is a general-purpose embedded database versus a search-specific format. See Persistence.
Feature surface
Meilisearch has shipped vector and hybrid search and geo search; Tachyon has not yet — both are tracked on the Roadmap. Meilisearch also offers managed hosting and enterprise clustering; Tachyon is self-hosted only, with no enterprise tier.
License
Meilisearch is MIT-licensed. Tachyon is Apache-2.0. Both are permissive with no copyleft obligations — the practical difference for most users is minor; Apache-2.0 additionally includes an explicit patent grant.
Choose Meilisearch when
- You want instant-search-as-you-type tuned well out of the box with minimal relevance configuration.
- You need vector/hybrid or geo search today.
- You want a managed cloud option or enterprise clustering support.
Choose Tachyon when
- You want relevance scoring you can reason about and tune explicitly (BM25 + per-field weights) rather than an opinionated rule pipeline.
- You want an Apache-2.0 license with an explicit patent grant.
- You want the smallest possible operational surface: one binary, one REST API, Prometheus metrics, nothing else.