add sqlite usage and cost monitoring

[KeremTurgutlu]

This PR adds sqlite based usage and cost monitoring.

Important Notes

• Total cost calculate by litellm doesn't consider web search requests, here is a relevant issue: [Bug]: Anthropic web search cost BerriAI/litellm#11422. So it might actually be a good idea to export and use the cost calculation utilities mentioned in (2).

• Web search requests are not included in litellm usage metrics when stream=True, created an issue for this: [Bug]: server_tool_use missing in usage when using stream=True BerriAI/litellm#16631 and a PR: include server_tool_use in streaming usage BerriAI/litellm#16826.

Litellm PR is not a blocker for this PR, once litellm team merges it, the following formula will give the total AI cost even with stream=True:

def total_cost(self:Usage, sc=0.01): return self.response_cost + sc * ifnone(self.web_search_requests, 0)

Fixes: #49

[jph00]

Currently this make fastlite a strict dependency AFAICT, but doesn't add it to settings.ini. I don't think we should make it a dep, since most people won't need this.

Probably just moving it to a separate module would suffice?

BTW, are we using cachy? If so, why are all the outputs changing? (And if not - we should be! :) )

[KeremTurgutlu]

Currently this make fastlite a strict dependency AFAICT, but doesn't add it to settings.ini. I don't think we should make it a dep, since most people won't need this.
Probably just moving it to a separate module would suffice?

Sounds good

BTW, are we using cachy? If so, why are all the outputs changing? (And if not - we should be! :) )

Yes, cachy is enabled. I just reverted my changes, re-ran the notebook and there are still changes but no new entries in cachy json file. Diffs are mainly from non-existing output (maybe those message outputs were left empty and weren't run in a previous commit), and in some cases they are from new line characters (maybe related to a rendering change).

Running the nb multiple times actually still produce few new line diffs for me. So, it appears to be not a problem with cachy but rather a flakiness with how things are rendered. Will have a deeper look at it.

[jph00]

Looks like we're nearly there! :)

[jph00]

Thanks! :)