Hey folks — hitting a snag I think others might be running into.
I'm using DeepSeek Chat for a lot of marketing automation research — comparing ESP features, digging into deliverability stats, or looking for specific Klaviyo vs. Mailchimp workflow differences. Sometimes, it just doesn't know the answer (which is fine!), but instead of saying that, it'll confidently fabricate details.
For example, I asked it about a **specific, newer SendGrid IP warmup schedule recommendation** from last quarter. It gave me a detailed, plausible-sounding table with daily sending limits... that turned out to be completely made up. The real advice from their docs was different. When I called it out, it apologized and said it doesn't have real-time data.
That's the core issue: **I'd rather it admit the gap upfront.**
What I've tried so far:
- Explicitly asking "If you don't know, please say so" — this works maybe 60% of the time.
- Giving it a context window like "Based on publicly available documentation up to mid-2023..." seems to help a little.
- When it does fabricate, I've been replying with "Source?" or "Can you confirm that's from the official docs?" — that usually triggers a correction.
Has anyone found a reliable prompt pattern or approach to get it to default to honesty about its knowledge limits, especially for recent or niche technical topics?
Maybe we can crowdsource a good pre-prompt or phrasing that works consistently. It would save a lot of double-checking!
Billy
Always A/B test.