No answers to questions

[gsuttie]

Within the data folder there are a number of pdfs but if i ask questions like the ones in the video they dont knwo the answer - how do i check this and how would i upload my own data for me to be able to ask it questions on say a pdf i uploaded myself?

[pamelafox]

In our experience, we've found that this model requires more prompting to decide to search the knowledge base. For example, with the sample data, I always say "Hello", and then once they respond, I ask "What is the whistleblower policy for Contoso Electronics?"

We need to experiment with more prompt engineering or perhaps few-shot examples to get it to more reliably search the knowledge base.

To upload your own data, assuming you used "azd up" for deploying from scratch:
add the files to the data/ folder, then rerun ./scripts/setup_intvect.ps1 or ./scripts/setup_invect.sh

[gsuttie]

Awesome thank you!

[Piesson]

I'm trying to find a better way to prompt an AI model to search the knowledge base more often or always. I've tried changing the tool_choice from "auto" to "required" and using {"type": "function", "function": {"name": "search"}}, but these either caused endless loops or didn't work at all. Do you have any insights or suggestions on how to handle this? I'd really appreciate your help!

[pamelafox]

@Piesson I haven't done much prompt engineering for this RAG app, but for others using function calling, I've found it helpful to provide few-shot examples. Here's an example of the format: https://github.com/Azure-Samples/rag-postgres-openai-python/blob/main/src/backend/fastapi_app/prompts/query_fewshots.json
It would need to match the format of the search tool used in this app.