There is Milvus artifact but not documentation

[lordofthejars]

Is there any reason to have the Quarkus Milvus integration but not being documented?

[jmartisk]

Yeah I could come up with some excuses like "I haven't had time", or "Nobody is using it anyway" or so. I don't know if anybody is using it with Quarkus, I haven't seen anybody. It just thinly wraps the upstream langchain4j implementation, so anything written in https://docs.langchain4j.dev/integrations/embedding-stores/milvus should be still applicable, we just add a configuration model on top. So yeah, we could certainly at least add a configuration reference page.

[lordofthejars]

Well, I am developing some demos, and I use PGVector and Chroma. Now, I want to try Milvus just to have all of them

[jmartisk]

Ok I'll try to write something up when I can. Generally, if it's for a demo, I assume you could use dev mode with dev services, in which case you should only need to specify quarkus.langchain4j.milvus.dimension with the vector dimension corresponding to what your embedding model produces. And everything else should wire up by itself. But you may very well find some issues, and if you do, please do let me know. Feedback from anybody else is useful before I continue to work on it, and I haven't yet received any.

[lordofthejars]

Sure. Tomorrow I will start playing with it.

[lordofthejars]

If you want, leave the issue open for some time, so if I find anything, I can add it here like the dimension property.

BTW, couldn't it be injected by calling the embedding.dimension() function automatically?

[jmartisk]

BTW, couldn't it be injected by calling the embedding.dimension() function automatically?

We've discussed this in the past, for in-process embedding models it wouldn't be too hard because the dimension is known, but with remote models, it is hard, because just from a string ID of the model, you can't deduce the dimension without actually calling the model, and the framework probably shouldn't do anything like that by itself.

[jmartisk]

Also we don't know for sure, if there are multiple embedding models in the app (with potentially different dimensions), which one you're going to use with the embedding store.