jaluma
commented
2 weeks ago This is very relative. Part of the difficulty of create a RAG application is that each domain and files have a different scope. In theory, the setup is well configured, but you will need to make adjustments to achieve better results. Here are some ideas I propose:
- Get more nodes in the context window.
- Enable hybrid_search in Qdrant.
- Change how the text is splitted.
- Change the model. Gemini2, Llama3, or Phi3 are working relatively well. With the version we have, we aim to have a more or less general version that doesn’t always have to work perfectly. I encourage you to share your experience or open a PR if you find a better solution to help the rest of the community :)
Looking for help in understanding how this works and what's going on.
I'm using the recommended Ollama configuration.
What did I do wrong that I'm getting these incorrect answers? Or what do I not understand about how PrivateGPT works and what it means about the 'provided context'?
Please see attached images.
Both Taco Bell and descriptions for the character Senna are on Page 2. As mentioned above, maybe I have the wrong idea of how this is supposed to work but these responses are not what I'm looking for.
This is from the source document.
I don't think this is the right approach but I looked in the database for what the data looks like and it's like this - using DB Browser for SQLite, exporting the .bin, using a converter.
Thank you.