When using an AI agent to look up information, we usually just dump an entire webpage, website, or directory into the model and let it "brute force" its way through to pick out a few potentially useful paragraphs.

This paper proposes a different approach:

1. Use a query retrieval system to push only the relevant paragraphs to the AI. 2. Consolidate and deduplicate the information from those sections.

This approach cuts costs and runs much faster.

This method solves the efficiency problem, but accuracy cannot be guaranteed. It relies entirely on the capability of the query system mentioned in the first point here.

Another way to improve data accuracy is to extract and transform the semantic layer before data entry. This does make writing data slower, but semantic data is different from transactional business data.

Semantic data doesn't necessarily need to be consumed by operational systems immediately. From a database perspective, it doesn't require OLTP characteristics. Therefore, a better approach might be to let this data undergo a period of processing and cleaning to become meaningful semantic information before making it available for subsequent retrieval.