Cross collection query?

[lguarda]

First this is awesome, then

I't looks like query are bound to collection, which i assume is on purpose, However, when using many open-source libraries and odd build systems (like CMake) or whatever comes in your mind, it would be great to query across multiple collection.

So is it something possible or even desirable? So then each chat could be bound to specific list of collection.
Or this is a dumb question and obviously, the best solution is to simply create one collection containing everything that you need?

[Davidyz]

1. For famous projects like cmake, I think it's a better idea to use a web search tool that supports searching from a specific domain. If you're using codecompanion.nvim, the built-in web_search tool can do this already. It's just you might need to supply it with a URL to the repo/documentation site. It'll be a less computationally intensive job for PCs to do.

2. For cross-project query, the best I can suggest for now is to handle this from the chat interface (like codecompanion). You can simply tell the LLM this project A make use of project B. It's indexed and you can search for its source code via @{vectorcode_toolbox}. Most of the LLM that I tried will run @{vectorcode_ls} at start, so chances are, it already knows what project there are and can already pick up the right project. In fact, this helped me a couple of times when I was working on the codecompanion extension.

But if you're looking for a way to handle cross-collection query within the CLI, it's not supported yet, mostly because I wasn't happy about the quality of rerankers (or rather, the ones that are small enough to run on my PC). The reranker issue is important because people may use different embedding models for different projects, and you can't just compare embeddings from different models. You'd have to get the raw text and use a reranker to sort the texts from different projects. If the reranker doesn't give a reasonable ranking, it'll be just wasting tokens for nothing.