Improve the Performance of RAG

[wwjCMP]

I am also using the smart composer plugin. In almost all cases, the results matched by RAG in the smart composer plugin are more relevant than those matched by this plugin.

[logancyang]

What chat and embedding models are you using? That plays a huge role.

I've been getting complete opposite conclusion from our power users. Not to dismiss your observation, but Copilot is the only Obsidian LLM RAG integration that leverages hybrid search rather than pure vector search, so with a comparable embedding model in another plugin, Copilot results can only be better.

[wwjCMP]

What chat and embedding models are you using? That plays a huge role.

I've been getting complete opposite conclusion from our power users. Not to dismiss your observation, but Copilot is the only Obsidian LLM RAG integration that leverages hybrid search rather than pure vector search, so with a comparable embedding model in another plugin, Copilot results can only be better.

I use ollama/bge-m3 for RAG in both plugins.

[wwjCMP]

[logancyang]

Many factors can influence this, like chunk size, similarity threshold and search strategy. That's why I created Plus mode to not let users think about that. It's hard to tell where your settings are not optimized.

First things first, is your copilot index complete? You can check by "Inspect copilot index by note paths" and see if your target docs are indexed at all.

[wwjCMP]

I manually reindexed several relevant files to ensure they can be retrieved, but the results returned by RAG are still not satisfactory. It should also be emphasized that, although no strict controlled experiments have been conducted (which is also difficult to do), in my long-term use, the relevance between the sources retrieved by smart composer's RAG and the questions is better.

[logancyang]

If you can provide controlled experiments with detailed side-by-side results, I'm more than happy to dig in with you. Otherwise it's not very productive making claims like this. And I'd like to add that both plugins are open-source. You are welcome to make a contribution.

[wwjCMP]

If you can provide controlled experiments with detailed side-by-side results, I'm more than happy to dig in with you. Otherwise it's not very productive making claims like this. And I'd like to add that both plugins are open-source. You are welcome to make a contribution.

Of course, what you said is completely correct. My capabilities limit me to sharing some of my personal usage experiences.