AI has always been about retrieving information but bigger problem is understanding the information and its meaning.
We all know about RAGs and how it makes possible to have more task-specific AI systems. Then there are things like rules.md, agents.md and other file-based prompting techniques have been the used for while now.
As useful it is to know about these as developer. But building a demo and a production-grade reliable working system are very different things. In production, alot of things matter.
AI agents forget and return giberrish responses. Commonly due to loss of information, conflicting rules etc, this is mainly due to context’s storage and its management.
The problem is not lack of knowledge. We clean data, which might might not original source but we still won’t get a good response from an AI agent if its context is messed up.
Originally published on Substack.