@VisiveAI Not any fixed number. But buffering adds roughly the full response generation plus one check call.
The check is usually another model call, so it costs its own prefill and a short generation.
Streaming and answer checks fight each other.
Once FastAPI sends the first token, the 200 status is already out. A groundedness check run after generation can't block a bad answer. It can only flag it.
Pick: buffer and check first (slower), or stream and send a correction.
@Nikita26rai It's because of data drifting. You can refresh your database or check your embedding models. The retrieval quality gets reduced by this drifting. Had to run the eval pipeline for measuring and making right decisions
Every agent platform is shipping long-term memory. The harder problem: what happens when a memory stops being true?
I'm building MemoryMesh to find out. Biggest lesson so far: fix stale facts when you write them, not when you read them.
@zerovv_ Songe turns conversations and corrections into training data, so memory lives in the weights, trained locally.
I'm building the other side: memory outside the model, where facts expire.
When a correction stops being true, how does the next dream session unlearn it?
@pranitttt It's for Shopify mainly for now. Planning to expand it for a large portion of d2c stores.
It fetches all your Shopify data and history once you login and link your Shopify store with it.
@pranitttt It's a solution for online sellers,
It predicts the "risk of churning" customers. And trigger automated winback flows by sending personalized discounts, it's linked to twilio and resend also.
Seller can customise the whole flow as well.