The most interesting thing about Typesense is a commercial decision rather than a technical one, and it deserves more attention than it gets. The dominant pattern in open source infrastructure is open core: the base product is free and genuinely open, and the things you need in production are not. Clustering, replication, authentication, encryption and the operational features that distinguish a demo from a deployment get reserved for a paid tier, and the open version exists mainly as a funnel. This pattern is so prevalent that most engineers now assume it, and evaluate open source products by working out which features are missing rather than whether the product is good.
Typesense does not do this. Clustering, vector search, typo tolerance and the full functional search capability are in the GPL-3.0 binary, with no feature gating. The commercial offering is managed hosting, which is an honest thing to sell: you are paying for someone else to run it, not for the software to stop being deliberately incomplete. I think this is the right model and I would like more projects to adopt it.
The documentation reflects the same practical orientation. It is written to get you to a working search implementation rather than to impress you, covering schema definition, indexing, querying, faceting, filtering and the typo tolerance behaviour that is the main reason to use a purpose-built search engine rather than a database LIKE query. The vector and hybrid search coverage is current and treated as a first-class capability, which matters because the interesting question in search right now is how to combine lexical and semantic retrieval sensibly. Pure vector search is worse than people expect at exact matching, product codes and proper nouns, and pure keyword search misses conceptual similarity.
Hybrid retrieval is the answer and the docs treat it as such rather than as a checkbox. The comparison page deserves specific credit, and this is rarer than it should be. Most vendor comparison content is an exercise in selecting the dimensions on which you win. Typesense's comparison is substantially honest, including about its own weaknesses, and the most important of those is stated plainly: Typesense is weak at CJK segmentation and non-Latin scripts.
That is a significant limitation. Chinese, Japanese and Korean text does not delimit words with spaces, so tokenisation requires language-specific segmentation, and a search engine that handles this badly will produce results that range from mildly wrong to unusable. Arabic and Thai have their own difficulties. Meilisearch, its closest competitor, supports considerably more languages natively.
If your product serves any of these markets, this is not a minor caveat to weigh against other factors, it is likely disqualifying, and I would rather a vendor tell me that than let me discover it after indexing. The licence is the counterweight to the no-gating praise, and the trade is a coherent one that you still need to think about. GPL-3.0 is strong copyleft, which is precisely what allows Typesense to put everything in the free binary without a competitor building a managed service on top of it and taking the market. Meilisearch, by contrast, is MIT, which is more permissive and correspondingly easier to get past a legal department with a copyleft policy.
For most companies running Typesense as a search backend behind their own application, GPL raises no practical issue. For a company distributing software, or one with a blanket policy, it is a conversation to have before you build rather than after. Neither position is wrong; they are different bets, and the licence you find acceptable should drive the choice as much as the feature comparison does. The gaps are the familiar ones.
There is no course, no learning path and no certification, just reference documentation, which suits the audience but means a developer who has never built search before will learn the API without learning search. And relevance tuning specifically is thinner than it should be relative to the configuration reference around it. Relevance is where search projects actually succeed or fail: getting results to come back is easy, getting the right result to come first requires understanding ranking, field weighting, boosting and the iterative process of evaluating against real queries that real users typed. The documentation explains the levers.
It does not really teach the practice of pulling them, and that practice is the job. The community size is a practical consideration too. Typesense has a smaller community than Elasticsearch or Meilisearch, which in day-to-day terms means that when you hit an unusual problem, the chance that someone has already written up the answer is lower. This is the ordinary cost of choosing a smaller project and it is worth factoring in honestly rather than pretending the licence and feature comparison are the whole decision.
Three point nine. An honestly licensed product with a genuinely admirable no-gating stance, practical documentation and unusual candour about its own limitations, marked down for the language weakness, the absence of a teaching layer and thin relevance guidance.