Vector-only RAG is not evidence

Most “enterprise RAG” systems are a vector index, a prompt, and a demo that cites the wrong PDF with confidence. That is not a knowledge layer. That is a search accident wearing a citation chip.

Lexical retrieval is boring on purpose. Clause IDs, exhibit labels, claim numbers, vendor names, and policy versions are exact. When an SIU analyst asks what the 2024 MSA says about subcontractors, the system should find §8.3, not a neighboring paragraph that shares embedding space with indemnification.

Vectors are useful after the exact path exists. Optional rerank can reorder a candidate set the lexical layer already constrained. Making cosine similarity the foundation inverts the failure mode: the system is most fluent when it is least inspectable.

Evidence-grade retrieval has four boring properties. The source is named. The passage is bound. The permission traveled with the document. The pack that produced the answer has a hash. If any of those are missing, you do not have a record. You have theater.

Knolo Knowledge Images are built as reviewable units: sources, ACLs, lexical indexes, optional rerank artifacts, hashes. That is the V4 surface you can mount today. Counsel can ask what is in the image. An operator can diff two versions. A model can be swapped without the knowledge plane collapsing. V5 is the runtime that makes the same artifact the agent’s source of truth.

This is why HIVE refuses to lead with “we’ll embed everything.” Embedding everything is how last year’s draft and this year’s executed agreement become neighbors. Deterministic first, then optional rerank if the eval says it earns its keep.

If your current system cannot show the image, the passage, and the permission for yesterday’s answer, it is not ready for operators who get deposed. Bootcamp is how we put a real corpus through that test in five days.