Honestly i am about to lose it with DeepSeek V4 Pro. I switched over because the pricing was better for my startup project here in London but man it is being so stubborn lately. I have this massive data cleanup task that needs to be finished by Friday morning or the whole launch is gonna be delayed and I just spent three hours trying to get it to follow a simple schema. It keeps hallucinating fields that arent even there or it just ignores my negative constraints entirely like its not even reading them.
Ive tried chain of thought and few-shot examples like I used with GPT but it just seems to get confused and starts looping the same three sentences over and over. I really need to make this work because my api budget is capped at 200 dollars this month and i cant afford to go back to the more expensive models right now but the output is basically garbage at this point. Its so frustrating when you know the model is supposed to be smart but it just wont cooperate with the way you talk to it. Does it need specific xml tags or something different for this version? What is the best prompt engineering strategy for DeepSeek V4 Pro to actually get it to behave?
Try wrapping your instructions in XML tags. I've found DeepSeek V3 67B API follows schemas much better when you use specific tag blocks so it dont loop.
> Try wrapping your instructions in XML tags. To add to the point above: spot on. Im satisfied with XML for DeepSeek V4 Pro but DeepSeek Coder V2 236B follows schemas better, tho its slower.
> It keeps hallucinating fields that arent even there Building on the earlier suggestion, you might want to consider dropping your temperature to exactly 0.0 for those data tasks. DeepSeek gets a bit too creative otherwise... also make sure to explicitly define the JSON keys in the system prompt. Be careful not to overload it with negative constraints tho, just focus on the schema. If it still fails, maybe try DeepSeek-V3 671B API as a fallback for high-precision logic.