Using Authorization header to specify Openai API key on a per-request basis instead of config.yaml

[Matteo-Frattaroli]

Hello,

first of all kudos: I recently started using memobase, it seems very well designed and it works really well.

For a project of mine, which uses the OSS version of your api server, I was wondering if you have in roadmap a support for specifying OpenAI api keys on a per-request basis.

Exploring the source code i noticed that you (rightly) instantiate a global openai client once and use it only when flushing the memory buffer (but correct me if I'm wrong). I'm just spitballing here, becuase maybe I'm missing somethig, but I was thinking that if, and only if, an Authorization header is specified, a new AsyncOpenAI instance could be created with the specified api key attribute and destroyed in the lifecycle of a single flush request otherwise falling back on the existing global instance.

To avoid excessive dependency injection, as i saw there is a pretty deep chain to go down to the actual request, maybe a context var can be used to store the actual instance to be used for a specific request without major impacts on the code base, performances or on the current functionalities.

If you find this proposal interesting I would be glad to contribute. Let me know what you think!

[gusye1234]

hmmm, can you elaborate more in which case you will need a per-request OpenAI API key? I guess you might have a lot of OpenAI Keys, and you want to rotate them for use?

[Matteo-Frattaroli]

Hi @gusye1234, I'm terribly sorry for the embarassing delay of this reply, I had to completely shift my attention from this topic due to personal matters.

Your guess was indeed correct. Having several customers, each one of them would have to use its own api key in order to get a correct billing.

If you'd be willing to reopen this issue (or let me know if you want me to open a new one) I'd be happy to finally implement this feature as proposed in my last message.