Track usage, add system prompt and timeout

[HerrIvan]

A few improvements to the openai client.

• Added a function to wrap the call to async generate_chat, s.t. it can process the usage data and add an additional parameter (the system prompt).

• Added the system_prompt parameter: not groundbreaking, but just required a small change. It may reduce the need to do prompt composition. There was issue about this was closed as not planned. there no 'system' role in the gpt-3.5 when build prompt? #246

• Added a timeout parameter: Without it sometimes the script was waiting for minutes and minutes for a 1-token response.

• Added the processing of usage: Accumulated usage is stored within each instance of the client. Return the number of prompt and generated tokens for API providers #387

[HerrIvan]

It may be that one is connected to different models (GPT-3, GPT-4, …) in a same application. Therefore better keep the counts separate. I don’t know if that answers the question or is it something different? El El vie, 24 nov 2023 a las 15:41, Rémi Louf ***@***.***> escribió:

…

***@***.**** commented on this pull request. ------------------------------ In outlines/models/openai.py <#388 (comment)> : > + def generate_chat( + self, prompt: Union[str, List[str]], config: OpenAIConfig + ) -> np.ndarray: + """Call the async function to generate a chat response and keeps track of usage data. + + Parameters + ---------- + prompt + A string used to prompt the model as user message + config + An instance of `OpenAIConfig`. + + """ + results, usage = generate_chat(prompt, self.system_prompt, self.client, config) + + self.last_usage = OpenAIUsage(**usage) I thought the idea of attaching the counts to the model class was good. Is there any reason why you changed your mind? — Reply to this email directly, view it on GitHub <#388 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A6ZVXMGZPGMVVM5C2U3A6YTYGCWY5AVCNFSM6AAAAAA7ZEZZ4WVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTONBYGA3DCNZUGM> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[HerrIvan]

I’ve created a new method because the “cachable” function was called from two different places, and that would’ve mean duplicating the code that keeps track of the token consumption. (I am afc right now) That being said, I am sure the solution can be improved :) El El vie, 24 nov 2023 a las 15:40, Rémi Louf ***@***.***> escribió:

…

***@***.**** commented on this pull request. ------------------------------ In outlines/models/openai.py <#388 (comment)> : > @@ -158,7 +182,7 @@ def __call__( ) ) if "gpt-" in self.config.model: - return generate_chat(prompt, self.client, config) + return self.generate_chat(prompt, config) What I would do instead is move the response + token count extraction here instead of creating a new method. — Reply to this email directly, view it on GitHub <#388 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A6ZVXMELHAP7DC76BA3RRY3YGCWWFAVCNFSM6AAAAAA7ZEZZ4WVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTONBYGA3DANRUG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[HerrIvan]

With the last version of this branch, the perscache serialization throws an error.

TypeError: Object of type CompletionUsage is not JSON serializable

I haven't found any better solution that converting the usage object to a dict before returning it.

Afaik, other solutions would imply modifying the serializer.

In any case, thanks for the work you are putting in this PR.

[rlouf]

I have updated the documentation and rebased the commits. Thank you for your contribution!