LLMs are statistical word association machines. Or tokens more accurately. So if you tell it to not make mistakes, it’ll likely weight the output towards having validation, checks, etc. It might still produce silly output saying no mistakes were made despite having bugs or logic errors. But LLMs are just a tool! So use them for what they’re good at and can actually do, not what they themselves claim they can do lol.

OpenWebUI connected tabbyUI’s OpenAI endpoint. I will try reducing temperature and seeing if that makes it more accurate.

Context was set to anywhere between 8k and 16k. It was responding in English properly, and then about halfway to 3/4s of the way through a response, it would start outputting tokens in either a foreign language (Russian/Chinese in the case of Qwen 2.5) or things that don’t make sense (random code snippets, improperly formatted text). Sometimes the text was repeating as well. But I thought that might have been a template problem, because it seemed to be answering the question twice.

Otherwise, all settings are the defaults.

I tried it with both Qwen 14b and Llama 3.1. Both were exl2 quants produced by bartowski.

Perplexica works. It can understand ollama and custom OpenAI providers.

Super useful guide. However after playing around with TabbyAPI, the responses from models quickly become jibberish, usually halfway through or towards the end. I’m using exl2 models off of HuggingFace, with Q4, Q6, and FP16 cache. Any tips? Also, how do I control context length on a per-model basis? max_seq_len in config.json?

Seems to be the only necessary thing in my case! Thanks.

Yeah I definitely have the default GTK chooser. Guess I have some config playing to do later.

Can you explain a bit more about this and how to configure it? When I use FF on gnome, the save dialogue just looks like other dialogues?

How much speed are you actually getting on Mixtral (I assume that’s the 8x7b). I have 64 GB of RAM and an AMD RX 6800 XT with 16 GB of VRAM. I get like 4 tokens per second with Q5_K_M quant.

A vector search converts your query into magic numbers, and then searches the database for other magic numbers that are “similar” (closet to it in the vector space, which is basically an N-dimensional graph of points and directions). These results are then returned as snippets to the LLM and it does stuff from that point.

The effectiveness of the vector search depends on how Open WebUI breaks up the documents into smaller sections, and how good the embeddings are.

I’m not exactly sure what you want to achieve, but you might have success in using an LLM to summarize the documents beforehand, using a specific prompt to extract the info you want, then feed that into the vector DB. This would require some scripting, of course.

The easiest thing to do is try it. See if Open WebUI’s vector search is able to handle it. Make sure to use a good embedding model like nomic-embed-text (can be found on ollama.com). You can change the vector search settings in the documents settings from the workspace on OpenWebUI.

Edit: https://ollama.com/library/nomic-embed-text

Open WebUI’s document management loads everything into a vector database. When you use the hashtag, it will trigger a search against the vector database based on your prompt. These results are run feed into the LLM. Open WebUI should generate a hashtag that can reference all the documents. But the quality of the results will be influenced by the embeddings and the LLM that responds to you.

Install ollama. It has ROCm support (on Linux at least). Then hook it up to your favorite client. It has its own API and an openai compatible one.

KoboldCPP has ban tokens that prevent those tokens from being output. Otherwise just put it in the prompt and it should probably work.

Doesn’t gnome already have this?

Cygwin ftw

I would think access to the training data, or at least no restrictions on what you can do with the model, would be a good definition.

Ah right. What I really meant to ask was if it can do protocols other than http.

Which I don’t think it can…

Are you able to tunnel ports other than 80 and 443 through Cloudflare?

But wouldn’t you calculate the time in the future in the right time zone and then store it back as UTC?