so I have been trying to get a LLM to interact with my journal notes in Obsidian (easy prompts like "what have I written about xyz")
first I used the Obsidian copilot plugin to link it up with gemini 2.5 flash-preview
I also tried GPT4all with a local model phi-3 mini instruct (4B parameters) and linked it up to my Obsidian vault

now the results are very wishywashy: the LLM gets very simple things right, but most of the time it doesnt use all relevant source entries or it uses completely irrelevant sources
it also isnt very precise, for example it finds the right source paragraph, extracts the right info, but then jumps one paragraph back to integrate irrelevant info into the answer

I have no idea if those free models just arent powerful enough or if I just need to finetune the model's parameters