How to restrict the response within pdf content?

[TarunKC261]

As shown in screenshot, when asked about prime minister it gave correct answer with incorrect citation. I want to restrict these type of questions which are outside the scope of pdf and show the response as "Question out of scope".

[Dag-Calafell-MCA]

You might try modifying the AOAI prompt to instruct it to only answer questions from the retrieved documents.

[TarunKC261]

I have added below instructions in the prompt but still it is giving results outside the scope of pdf uploaded. """Answer ONLY with the facts listed in the list of sources below. If there isn't enough information below, say you don't know. Do not generate answers that don't use the sources below"""

Also added the instructions multiple times """Strictly generate the answer only from the sources listed below. If there isn't enough information below, say you don't know. Do not generate response outside the content below."""

[Dag-Calafell-MCA]

After deploying, I noticed the "retrieval mode" in the "Developer Settings"... you should try "Vectors only" because this should limit it to just the PDFs (vector db). I haven't reviewed the code however.... but I did test "What are the federal laws which govern Worker's Compensation?" and it seemed consistent with my assumption.

[pamelafox]

Hm, I don't think that would affect whether ChatGPT decides to answer outside of the indexed data, that just affects whether it does both a semantic and vector search (see https://techcommunity.microsoft.com/t5/azure-ai-services-blog/announcing-vector-search-in-azure-cognitive-search-public/ba-p/3872868)

I think that ChatGPT just isn't always adhering to the prompt. It may be worth trying GPT-4, if you have access to that, as I have heard it's better at sticking to instructions.

[Dag-Calafell-MCA]

Really good insight there, so Cognitive Search can be my vector database now. I did not know!

[itmilos]

@pamelafox @Dag-Calafell-MCA During Semantic Kernel Office Hours, this was discussed. It appears that currently, only an effective system prompt can prevent it.

I'm considering using an approach similar to what the HackAPrompt team utilized.

ref. https://huggingface.co/spaces/jerpint-org/hackaprompt/blob/main/hackaprompt/utils.py

Also good reference for how to HackAPrompt https://github.com/terjanq/hack-a-prompt/tree/master

[github-actions[bot]]

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this issue will be closed.

[github-actions[bot]]

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this issue will be closed.