I'm of the understanding that the way that the Research Assistant selectively chooses to summarise parts of files to send to models is a holdover from back when there was no per-user token billing, and Kagi were literally hoping users would run under fair use limits. From this perspective, I can understand why this was added in the early stages of assistant.
Now, however, I'm paying for Ultimate. I am paying for tokens and I don't even come close to using all of them every month - I don't really like even getting close to the final quarter, to be honest. I have credits to spare. And this "summarising key details" framework is making LLMs absolutely unusable through assistant for key use cases for me. Just like in this example where I'm trying to request fixes to a graphic to be used in OBS:
This response alone is just absurd. I passed a html file to the model hoping that the model would be able to edit the html file. In exchange, it's of my belief that Kagi has actually created a markdown summary of the visible contents of the file and passed it to the model. In the same way, if I pass it text files or PDFs, it doesn't actually give the model the PDF - models have gotten really, really good at PDF reading since the feature was added - it just creates Kagi's executive summary of what it thinks is important in the document and passes it on.
In short, no thanks. I want to use Assistant to use LLM models and their associated features. I'd like to do things with the features LLMs have. I don't want Kagi completely opaquely sitting in the middle and deciding what it thinks is important enough to send. This actually seems like an incredibly anti-user design pattern - I don't know what it's doing, I don't know what it's summarising, all I know is that my context will get destroyed and the job won't get done - I may as well downgrade to Pro because Ultimate's main feature just doesn't work randomly.
Please let users - at least Ultimate users - disable the Research Assistant's summaries of files. Much like how I can disable web searching if I don't need it to save the time and tokens, I can disable this if I'm happier using more cost, for sure, to actually use the models properly.
(note: this post is a duplicate of the same frustration but can be deleted - this is a more comprehensive explanation of the issue. apparently i'm not allowed to delete my own posts Mod edit: Deleted 👍️ )
