Hoping to conduct a POC on the cloud version soon. Just need to verify basic stuff first.

running into a bug on 30rc27 where i

best_field_weight

and textmatch doesnt reflect the best field, but a lower-weighted field

Hi, I tried upgrading our cluster to v29 around 4 hours ago, but it appears to be stuck. The bigger issue is that sometimes when we query a collection, the data is incomplete, only 10% of the entries exist in the collection. When I triggered the upgrade, I noticed Node 1 cycled through healthy -> unhealthy a few times, but then it remained healthy permanently. I found another thread here with a similar issue, where apparently they were doing many updates instead of bulk import. in our case we do run some scheduled updates, but we use the bulk import feature. We're on the developer support plan if important

Hi everyone! Has anyone been able to successfully host Typesense on Azure Container Apps (though I suspect my issues aren't specific to that)? Whenever Azure tries to spin up another replica, I get a few errors, most notably: "It seems like the data directory /data/meta is already being used by another Typesense server." I suspect it's because of some locking mechanism? I have mounted a persistent volume (using Azure Files w/SMB) that is shared between replicas, but not sure how to resolve it. I want the data to be shared between them 🤔

Hi! I have a set of externally-generated vectors, and I want to find similar documents. I found this in the documentation:

{
  "vector_query": "embedding:([], queries:[smart phone, apple ipad], query_weights:[0.9, 0.1])"
}

1. Is there a way to do the same thing, but if I have externally-generated vectors? 2. If yes, would it do a nearest neighbour for each vector, or does it just average the vectors according to the weights and run a single nearest neighbour for that average vector?

Hi! I am getting some unexpected results. I was expecting both following queries to return exactly the same results: Hybrid search but with alpha=1 (which should basically sort by embedding search alone)

"query_by": "name",
            "group_by": "productId",
            "group_limit": 1,
            "num_typos": 0,
            "vector_query": (f"embeddings:([{embeddings}], k:100, alpha:1)"),
            "q": search_query,
            "per_page": 10,

vs (semantic search alone)

"group_by": "productId",
            "group_limit": 1,
            "num_typos": 0,
            "vector_query": (f"embeddings:([{embeddings}], k:100, alpha:1)"),
            "q": "*",
            "per_page": 10,

But these are producing very different results. The first one is still definitely influenced by keyword search. Even when I tweak the

alpha

parameter, I seem to always get the same results. Could you help me understand these differences? cc @Joao Neves

Hi! I have a collection on a multi node cluster. will adding the faceting make the collection unavailable?

Hey all, I have a query question. Basically, I am working with like a logic builder, but I need to disable segments of logic in Typesense. In SQL, I can do that like this

Select * from a WHERE a.b = 'correct' AND 1=1 AND a.c = 'correct';

The 1=1 lets me basically drop clauses without worrying about the conjoining OR / AND, whether we’re leaving parens or not, etc. In Typesense, is there a similar way to “skip” a clause by converting it into something like:

filter_by: 'b:=correct && 1=1 && c:=correct'

Hei quick question what is the recommended way to go these days, if we would like to build a semantic search ? Should we do a Vector search and the embeddings our self or should we go with Natural Language Search way https://typesense.org/docs/29.0/api/natural-language-search.html#use-case ? If we do the embeddings our selves will this give us some kind of and endge to build out the recommendations as well like here: https://typesense.org/docs/guide/recommendations.html#step-6-generate-recommendations

hi guys, can I use the embedder model that I created on the local ollama instead of using the typesense models in huggingface as the embedding model in the vector search section? In this link, in the auto-embedding section, it seems I cannot add an ollama model to the model_config value.

Hello, I am receiving some unexpected results when querying with vs. without group_by, as well as using a filter in group_by. Hoping someone can shed some light on this behviour. Running Typesense 29.0 in Docker Compose Case 1:

"page": 1,
    "per_page": 50,
    "include_fields": ["owner_article_number", "id", "virtual_id"],
    "q": "HH",
    "query_by": "owner_article_number",

Returns a document with

id: "1412145", virtual_id: 281343, owner_article_number: "HH1500038"

(Total 10 hits) Case 2 - as Case 1 plus:

...
    "group_by": "virtual_id",
    "group_limit": 99

Does not return any group for

virtual_id: 281343

, but 4 other groups with a total of 10 documents (Note, same # as Case 1, but some differences in documents returned) Case 3 - as Case 2 plus:

...     
    "filter_by": "virtual_id:281343"

Returns a group for

virtual_id: 281343

exactly as expected

am I correct in assuming you still cannot point one alias to another alias?

Hello! Do synonyms apply to vector queries, or are they intended only to apply to keyword queries? We'd like to use synonyms to expand acronyms into their full definitions when using vector search, but they only appear to apply to our keyword queries.

Does there exists any migration tool for typesense databases?

Hi Is there a GUI of some sort I can use against a docker instance? I'n trying out this product and a GUI would help

Another question - I'm working on a system which is highly active, and may create millions of documents in a few minutes (during peak hours). When I tried simulating this on a local docker instance, request fail quite fast with 503. What controls the ingest rate? is it a config? RAM? both?

Hello! Sorry if this is asked before, but I am trying out v30 rc29, and the rc docs talks about the new endpoints for synonyms, like synonym_sets, but for some reason they do not respond. But the old endpoints responds just fine, but feels like they do not get applied in search (have not done much experiments here so this might be invalid). But is it me doing something wrong, or is the new endpoints not enabled yet in rc29? Code is merged a while back for new synonyms so looks like it should.

❯ curl -s -w "\nHTTP Status: %{http_code}\n" \
  -H "X-TYPESENSE-API-KEY: motoaction-typesense-dev-key" \
  -H "Content-Type: application/json" \
  "<http://localhost:8108/synonym_sets>"

{ "message": "Not Found"}
HTTP Status: 404

for example. While the old endpoints do respond:

curl -s -w "\nHTTP Status: %{http_code}\n" \
  -H "X-TYPESENSE-API-KEY: motoaction-typesense-dev-key" \
  -H "Content-Type: application/json" \
  "<http://localhost:8108/collections/products_se/synonyms>"

{"synonyms":[{"id":"synonym-3542ccf1","root" ......

curl -s -w "\nHTTP Status: %{http_code}\n" -H "X-TYPESENSE-API-KEY: motoaction-typesense-dev-key" -H "Content-Type: application/json" "<http://localhost:8108/debug>" 
{"state":1,"version":"30.0.rc29"}
HTTP Status: 200

Also trying "by the book" (RC docs) to create one in case list would not work when there is no sets:

curl "<http://localhost:8108/synonym_sets/clothing-synonyms>" -X PUT -H "Content-Type: application/json" -H "X-TYPESENSE-API-KEY: motoaction-typesense-dev-key" -d '{
  "items": [
    {
      "id": "coat-synonyms",
      "synonyms": ["blazer", "coat", "jacket"]
    }
  ]
}'

{ "message": "Not Found"}

From the typesense docs:

Typesense is an in-memory data store. When you restart the Typesense process, we read the data that was previously sent into Typesense (and stored on disk as a backup) and use it to re-build the in-memory indices. The amount of time this takes depends on the size of your dataset. So until your data is fully re-indexed in RAM, the node will return an HTTP 503.

What is a typical amount of time? For 12M records (about 8GB of RAM) I am seeing 15 minutes. Is that typical or do I likely have something misconfigured?

I am on Typesense 26. I have a query that returns records when I filter by tenantId. One of those records has a first_name "Annie", but when I search filtering by tenant id and q=Annie no results are returned. Why might this be? Are there known issues with filter_by and query with a large number of records in TS 26? I have tried setting max_candidates to 100000000. I am wondering if there is some issue where typesense finds other results and then filters them out after with the filter_by tenantId and so shows nothing.

It seems almost like I have these "ghost records", they do not come up no matter what, except when listing with no query. They don't come up with an exact filter or a search query

Hi team 👋 We’re doing bulk writes to our Typesense cluster using the

/import

endpoint (sending

.jsonl

documents). In our case, we only send delta or updated documents during each sync operation — not the full dataset. Our goal is to ensure that the collection in Typesense exactly matches the current dataset we have locally. So, we’d like the documents that are not part of the latest bulk write to be automatically deleted from the collection. Is there a recommended or built-in way to achieve this behavior? For example, does Typesense support something like a “replace all” or “sync” mode for imports — where it automatically removes documents not present in the latest batch? Or would we need to handle this manually (e.g., track IDs and delete them ourselves)? Thanks in advance 🙏 cc @Sahil Rally @Atishay Jain

from a quick search i couldn't find a windows binary for typesense, only through WSL. is there a way to have typesense bundled in a local windows c# app?

I am following the steps here to upgrade typesense from 26 to 29. I have followed the steps exactly, and upgraded typesense on a single follower node (state 4) and restarted that follower. After the restart, I pinged the health endpoint on the newly upgraded node and it is now the leader (state 1) and the other two are followers. Is it expected that they would change over without restarting the leader?

Hi! Getting started with typesense and ran into an issue with filtering array of objects. Do I understand correctly that this query should produce result matching both conditions? filter_by=facets.id=68f237e6fb82d04dade59fc8&&facets.id=6909c02adce759581938cac3 My problem is that this query produces an OR result. Any suggestions are much appreciated 🙏 What I want to achieve after this simple test is something like: filter_by=facets.id=[val1,val3,val5]&&facets.id=[val1,val4,val6]&&facets.id:=[val4,val1] -> this should return only documents that have val1

Hi! I was wondering if it's possible to use scoped API keys with multisearch queries, applying different embedded params to different collections. I'm using Typesense 29

Hey Guys! we are running a typesense-cluster self-hosted, works like a charm 💪 We installed Typesense as a Service on a Server which runs multiple cluster-services (so basically we have 3 servers with multiple cluster-services installed) .. now we want to split them up & give every service separated servers we have now our old cluster and three new servers, my idea was • add one node to the cluster (as a follower) • wait till the node is fully synced • add the other two nodes like above • then make one of the new nodes the leader • change the loadbalancer • remove the old nodes one by one .. is there anything i should/must consider in that particular way? Would it be a problem if the old nodes are 28.0 and the new nodes have already 29.0 oder should we play save and do the upgrade first?

nts

Hey Folks - just hit this in our app and wanted to feedback - it would be great if the list of special characters was provided in the docs here: https://typesense.org/docs/guide/tips-for-filtering.html#escaping-special-characters

Is there an ETA on Typesense 30.0 release?