#if you don t include previous responses

6 messages · Page 1 of 1 (latest)

tacit blade
#

So do you think use previous responses in the system message and in the prompt it self? (its for text gen task, so can be long prompt)

#

currently the process is:

[{'role': 'system', 'content': 'previous responses'}]

prompt: 'Previously, you wrote: {previous responses}. Continue from your last response and.... '

upbeat fable
#

Previous messages from assistant role should still be called that. You might need to estimate the token size and start dropping content from the history like ChatGPT does for longer conversations.

#

Overall, your use-case might become more achievable once they provide access to 32K token prompts.

tacit blade
#

Yes, regrading the access to the 32K you totally right. In terms of dropping content, I'm load only the last response and not all of them. Just wondering if use in both of them is relevant or i can drop the previous response text in the prompt and just write: "based on your previous response, now write XYZ"