feat(openai): add tags & metadata at the call level

[Dam-Buty]

This adds the possibility to add tags/metadata to the OpenAI instrumentation, at the call level. The original feature (at the instrumentation level) is retained.

This is meant to be used like so :

const instrumentedOpenAi = client.instrumentation.openai();

await instrumentedOpenAi.chat.completions.create(
  {
    model: 'gpt-3.5-turbo',
    messages: [
      { role: 'system', content: 'You are a helpful assistant.' },
      { role: 'user', content: 'What is the capital of Canada?' }
    ]
  },
  {
    literalaiTags: ['tag3', 'tag4'],
    literalaiMetadata: { otherKey: 'otherValue' }
  }
);

[linear]

ENG-1685 add metadata support to generations

[desaxce]

We should have an afterAll that deletes the tag. Le pendant de client.instrumentation.openai({ tags: [testId] });.

[Dam-Buty]

I'm not sure i understand. Do you mean an option to remove the tags at some point so that all calls from now on no longer have them ?

In this case i'd rather recommend to either create a new client (there is no overhead, the client is just an instance of a class), or to tag at the call level. WDYT ?

[desaxce]

It looks like super high-level tagging. I'd only use it for A/B testing my code, using it with commit hashes for instance.

[desaxce]

I understand now that it's the old style...

[Dam-Buty]

Yes, there are 2 levels of tagging :

• at the instrumentation level, with this one you can use the usual OpenAI client
• at the call level, with this one you have to use the instrumentedOpenAi client

I left the high level one because otherwise it would break backwards compat.

[desaxce]

I would rather have one line that does literalaiClient.instrumentOpenAI() and then use the usual OpenAI client, to not have to change my application code.

[Dam-Buty]

Me too ! The issue is that this method adds parameters to OpenAI calls (literalaiTags, literalaiMetadata) and the only way for Typescript to understand them is to return this specifically typed client.

As i commented above, for the moment this leaves two options : you can use the original openai client if you don't tag or if you tag at the instrumentation level - but you must use this instrumentedOpenAi client if you tag at the call level.

[desaxce]

That's reasonable.

[desaxce]

Can we already add run and thread tags?

[Dam-Buty]

yes you can just add them to the constructor !

[desaxce]

Okay, that's what most people need.