We've had several issues working with upstream AI SDKs over the last few weeks. Over and above that, we're not really using the existing SDKs for their original purpose of making requests and receiving responses, but instead for high volume parsing and relaying of requests & responses - in which we've found several performance-related issues.
The APIs that we're interacting with are not that complex, so we should instead just implement the thin slice of the overall APIs that we actually interact with in a way that's ergonomic for us.
We should consider the ongoing maintenance and correctness trade-offs.
We've had several issues working with upstream AI SDKs over the last few weeks. Over and above that, we're not really using the existing SDKs for their original purpose of making requests and receiving responses, but instead for high volume parsing and relaying of requests & responses - in which we've found several performance-related issues.
The APIs that we're interacting with are not that complex, so we should instead just implement the thin slice of the overall APIs that we actually interact with in a way that's ergonomic for us.
We should consider the ongoing maintenance and correctness trade-offs.