Closed unicomp21 closed 5 days ago
I realize appeal of current approach is lower complexity, perhaps this means we should have two api's? The current approach, and then another secondary api/approach which uses sha hashes to store/lookup the cached prompts? ie If sha lookup fails then http error returns, and prefix/prompt needs to be reloaded?
@alexalbertt what does completed mean? Will we have a bandwidth efficient mechanism like evalsha?
The current approach will waste massive amounts of bandwidth? Those in bandwidth constrained environments (ie starlink) will be at a disadvantage due to the wasted bandwidth? Why not follow the redis evalsha pattern to eliminate redundant payloads and wasted bandwidth?
https://github.com/anthropics/anthropic-sdk-typescript/issues/516
https://redis.io/docs/latest/commands/evalsha/