habbywall wrote:
Is this anything more than just uploading the document to openAI and passing prompts through?Yes, the problem with GPT is that it hallucinates. So, to avoid this, I force it to use your document as its knowledge base. But there is no secret sauce here; the approach is straightforward.
I have set up a vector database (Pinecone). Then, I take your document and make it into smaller chunks by applying a sliding window. After that, I convert the text into vector embeddings and send the result to Pinecone, using vector similarity search when you ask your question.
The only prompt I have added is: 'Answer only if the answer is in the document, but give all the information you can:'
I had a server issue but now it runs without any problems. It uploads 5-6 pages/second so please try it out and let me know what you think!