It works well for retrieval: “Find my gas bill from January.”

Though even there similarity can be brittle. If I don’t specify the year, retrieval may surface the wrong January because multiple documents look nearly identical.

It really breaks down for aggregation: “How much did I spend on electricity and gas last year?” “Which months had the highest energy costs?”

At that point the problem feels misaligned with similarity search itself. You don’t want relevant chunks, you want structured values aggregated across documents.

Curious how people solve this. SQL tools? Structured extraction? Different agent patterns?

Comments URL: https://news.ycombinator.com/item?id=47922550

Points: 1

# Comments: 0