Hey everyone! I've been absolutely loving Kimi for research — that long context is a game-changer for my deep dives. But I’ve hit a snag the past few days, and I’m wondering if anyone else has seen this.
I uploaded a pretty dense, text-heavy PDF (a technical whitepaper). When I ask Kimi specific questions about the content, it gets the *general* topic right, but then starts inventing details. Like, it’ll cite a study that isn’t in the document, or attribute a quote to the wrong author from within the file. It’s not just a little off; it’s confidently making things up based on the overall theme.
My usual workflow is: upload, then ask it to summarize sections or pull out specific data points. This has worked flawlessly before! What’s different now? Is it the way I’m phrasing my questions? Could it be the PDF format itself? I made sure it’s searchable text, not a scan.
What about trying to “anchor” it more in the text? I’ve tried saying “based on the uploaded document…” but the hallucinations still sneak in. Has anyone found a prompt pattern or a setup trick that keeps it more faithful to the actual uploaded content? Really curious if this is a known quirk with a workaround.
Words matter