Use messages API and system prompt #7
Conversation
Thanks to @9j7axvsLuF for some sample code in the original issue tomviner#6
| # Generate a list of message dictionaries based on conversation history | ||
| messages = [] | ||
| current_system = None | ||
| if conversation is not None: | ||
| for prev_response in conversation.responses: | ||
| if (prev_response.prompt.system and prev_response.prompt.system != current_system): | ||
| current_system = prev_response.prompt.system | ||
| messages.append({"role": "user", "content": prev_response.prompt.prompt}) | ||
| messages.append({"role": "assistant", "content": prev_response.text()}) | ||
| if prompt.system and prompt.system != current_system: | ||
| current_system = prompt.system | ||
| messages.append({"role": "user", "content": prompt.prompt}) | ||
|
|
||
| return messages, current_system |
There was a problem hiding this comment.
I followed Simon's OpenAI plugin here. I don't think Anthropic allows mid-conversation system prompts, so I went with the most recent system prompt overriding everything else. I could imagine other behaviour, including generating some warning to the user if there are multiple system prompts
|
Just wanted to give a small bump here since about a week ago Anthropic announced that their messages API is out of beta. (No particular rush on this from our end in that we're fine to use our fork) |
|
Just flagging that this PR is now out of date since Anthropic has released v3 of their models. I can make another PR once they release haiku, though seems like this repo is deprioritized? |
|
Just found this PR while trying to add new models on my own (v3 and v3.5). Looks like Claude discontinued claude-1-instant, while claude-2 still works. @tomviner how's things from your end, are you accepting PRs? |
|
I don't think this is maintained. Rather use: |
That's right, I'll add a note to the README |
Thanks to @9j7axvsLuF for some sample code in the original issue #6
I had two main goals:
I didn't use the code from @9j7axvsLuF verbatim because I found that it broke the API. A tertiary goal I discovered I had was to maintain the API so that I could swap between models seamlessly. I'm not too in the weeds here, but can we create a conversation object to prime the conversation and "put words in Claude's mouth"?
Apologies it's all lumped in one big commit/PR, but it's a pretty small update overall