I’m testing out assistants for an app idea I had, but I’m struggling to have it cite multiple sources even when prompted. Instead it always seems to cite a single source, multiple times.
I’ve been testing with some copy/pasted articles off the internet.
I’ve uploaded 2 files, both about a recent story involving Derek Chauvin.
Instructions: You are journalist gpt. You summarize news into single paragraph briefs and cite your sources.
Model: gpt-4-1106-preview
Retrieval (check) - 2 articles (I could copy the text if that is helpful but I uploaded them both and explictly hit “save” on the UI)
Prompt: “Summarize recent articles about Derek Chauvin. Cite multiple sources if possible”
Have others had any luck getting gpt to cite multiple sources? Any model
Thanks!
Update: I peaked into the API response via the logs and I do see them both in there actually - i guess it just doesn’t always print them out!
Hi Evan - Just a thought from a non-technical perspective if still relevant: I would also expand your instructions and clarify that you expect the assistant to draw on multiple files for the task, you may even want to explicitly reference the file names in the instructions. I know you do this with the prompt already but the instructions can have quite a material impact on how your assistant executes tasks.
Nice, I will clarify that. It actually does seem to draw on multiple files, it’s just not clear where in the files it’s drawing from. Will post another question soon haha
If by “multiple sources”, you mean for the bot to cite sources not present in the two uploaded files, you might try accessing the web from the chat session. The LLM by itself is not likely to provide any sources. I also tried Tavily, and this does a good job by using web search + GPT summarization.
