OpenSource like Llamaindex vs Azure... clients always asks about security... I'm trying to find the best opensource security architecture with Llamaindex.
Hey! Today a client told me that with the chatGPT's they have "build" they are having worst results than months ago. Thinking about it, having ChatGPT4 since then, I'm guessing that when they build the GPT, the text-embedding-ada-002 and now they upload files to compare against this GPT. Could it be that the text-embedding-3-large with persent file doesn't talk properly with VectorDB already embedded in the past? I've told them to have the files deleted and uploaded again. Guessing that it will be better. Of course I've told them that with solution like Llamaindex choosing the embedding and having the same model being use, this it wouldn't happen. Any thoughts on this?
@kapa.ai Hello. I'm raising this question/thought so maybe someone could answer. I know it's no easy to answer. Based on my knowledge about models, I'm seeing that the bigger and greater models, the better. But I don't understand why there isn't a simplify model that knows only how to talk and then teach it on a field, so a simple and light model can answer better than GPT4 (the best out there now). How would be the best training method to transform a light model so it can answer better giving it some RAG files and context? Is this possible? How? On my learning path, I've found MEOW app plugin for WordPress for fine-tune a model, and the main problem is the cost for training + the cost of inbound and outbound afterwards (fine-tuned models inbound and outbound are way expensive than the main model out of the box, check it, it's soooo expensive). Also I'm thinking of Ollama + Llama3 + fine-tuned model in local, but I'd like to know how to have the model trained so the questions and answers are coherent. Anyone experience on this would be much appreciate. Thank you.
Hello everyone. My head is turning around (like the fools garden song from lemon tree) Because I'm seeking for the perfect RAG and I read about reranking and then about CRAG and then about Knowledge graphs, and then about text to SQL queries,... As far as I'm seeing, KG are the best and more accurate solution, and correct me if I'm wrong please it can't be mixed with Vector Databases, right. What about reranking. Please share your thoughts on your fundings and tests. I'll as soon as I do more tests. At the moment, Vector DB and reranking is what I'm using but I'm thinking on jumping towards KG.
Hello, I have an important doubt, even ChatGPT it looks can't solve. While everybody is talking about RAG's, and training them over GPT 3.5, Llama2, etc... Nobody talks about that this models are filled of information I don't want for the specific RAG training. If I want a RAG in Medicine or whatever specific field, I have to build the RAG on top of a model that has info about cars, planes, lettuce, shoes... that I don't care about it. Is there an existence of a model that "only" knows how to talk and reason without all this info, so it's a virgin LLM ready to be trained in the specific field?
