Spaces:

huggingchat

/

chat-ui

like 1.11k

Running

This is what the Huggingchat needs.

Looks like the models that support tools work with web searching directly, they (surprisingly) can make not one, but sometimes two search queries, with better understanding of the context. That's great, but still needs to be improved.

What about the models that don't support tools? I read somewhere on the community tab, that web searches are handled by the same model (as i remember Mistral or something), which also generates conversation topics. It does it's job, but much, much worse, starting by the fact that it's limited by one search query per main model's response. Second thing is the way how it actually works: I wanted to look up for some old mobile phones, and gave a prompt "Sony Ericsson S003 specs" to the Deepseek-R1. I also set a domain search to the "gsmarena.com". Mistral didn't make an efficient search query, didn't choose good sources, and even managed to confuse R1.

Seems like the problem was not only in sources chosen, but in the way what information was handed to the R1. Mistral decided that choosing "compare sony ericsson s003 and sony ericsson aino" as a source and convincing the main model that Sony Ericsson Aino specs listed in the comparison are actually a Sony Ericsson S003 specs is very funny. But R1 isn't as dumb as it looks like, and instantly spotted that information from sources differs, but sadly, it still affected the quality of a final response.

If all these issues will be solved, Huggingchat may even become a Perplexity competitor. At first, all models should have direct access to the web search. Also, for the whole process using web search, they should use a different system prompt (not one used for making responses), have access to some part of the previous conversation for context (this is necessary to avoid token overflow; also, this feature must be customisable, as it may confuse the model), and, of course, to the current user's message. User must be able to set this system prompt and choose how many search queries will be made. Now, all these changes to the current web searching logic may greatly increase the quality of search results. So now, that's the second part – way how the model will parse search results. Once again, minimum and maximum amounts of sources to be parsed can be set by the user. Most important part is that the model must select related information on the page based on it's source code. As said before, Mistral ultimately fails at parsing page contents. Also, model must be able to follow links to other pages to better expand on the topic. This feature is toggleable too. And the last thing to do – use parsed information to generate a response, that's pretty simple, doesn't need to be changed.

By the way, developers, please remove these dumb limitations in the assistant creation window! Why it's not possible to set model parameters to 0, and why assistant requires to have at least 1 character as the system prompt?

And another thing I wanted to ask for – will 72b version of Deepseek be on the Huggingchat? This question has already been asked, but nobody answered