Thank you. I recently made one inspired by this but more around business check it out if you're interested kzfaq.info/get/bejne/ZqeWhLt3z5_RZ40.html

Thank you

Glad it was helpful!

Glad it helped!

Thank you

Do you want to reduce spend? Do you want to maintain confidentiality? If so then yes.

Thanks brother 🙏🏽

Thanks for sharing

When you use a vectorDB you often create vectors of chunks of information using an embedding engine then store that information into the vector database (i.e. Pinecone). Then when you ask a question you chunk that question up in the same way you did with the information you have stored in the vector database using the same embedding engine. That question gets scored with those vectors and a mathematical comparison of the question chunks position vs. the closest K results of information chunks is retrieved. The retrieval IS the information chunks. Those similar information chunks that are semantically similar to the question then usually get fed along with that question to ChatGPT or some other LLM to provide an answer. Hallucinations happen because the LLM doesn't know what it is talking about. It only knows that things are similar and are more often found near each other when examining billions of parameters along many dimensions. The results appear coherent and are often right but being right or wrong is a side effect of the information being semantically similar and the probability that the desired result will be commonly found within the parameters located near that data. Example you ask a question about how cats get around. Most of the examples semantically similar to that will probably locate information in line with cats walking on their legs. But the information is randomized against a seed when it is retrieved and even uncommon results do have some chance of being chosen. So some essay a five year old wrote about their flying cat could be chosen as the basis for the completion even though semantically it's unlikely that this chunk of information will be chosen. Just to be clear vector databases are more for information retrieval to feed into LLMs vs. a reason why an LLM may choose to hallucinate although if the K results fed in from a vector database query happen to all be irrelevant than ChatGPT or whatever LLM is being used may choose to use those vector DB references and produce something that doesn't give you the answer you want. But this is not the cause of the phenomenon in the majority of LLM queries as they are not leveraging custom vector databases as part of their usage.

@@comfixit Embedding engines/algorithms are essential parts and there are lots of choices. Does a vector DB provide a default embedding engine implementation or just provide an interface to hook in any embedding engines? If embedding engines are not part of vector DBs, it sounds to me there will be lots of interactions between a vector DB and embedding engines it uses since each image/doc must be translated to embeddings.

Yep for example chroma db

pgvector is open source that you can add postgre for vector support. Redis also has support (for in memory) if you need another layer so you can retrieve faster.

You mean in the title? Should I remove that ?

Done! Removed new. Is it fine now?

@@1littlecoder Well, it was "new" to me. This is the kind of pertinent info, explained in understandable terms that I have come to expect from you here. As more serious AI dev's become disenchanted with the other channels that waste time with their sensational click-bait titles - which offer nothing of usable substance - your channel and reputation as a real expert will grow. It will take time, just keep doing what you're doing!

@@1littlecoder that's better 👍