Type your evidence.
Embeddings give you “similar.” Type tags give you “relevant.”
Why it matters
When a RAG system retrieves “the most similar passage,” it has no idea what kind of passage that is. A forward‑looking statement and a historical fact about the same number look almost identical to an embedding model. So do an opinion and a policy. The retriever returns whichever scored higher; the LLM treats them all the same way; your answer ends up grounded in the wrong type of evidence.
The cost of not doing it
You ask “What are our compliance obligations under GDPR?” The retriever returns a chunk that’s an internal opinion piece from a 2019 blog post titled “GDPR compliance is mostly common sense.” Both look relevant. The opinion scores higher because it’s more conversational. Your AI answers based on the opinion.
How BRAG does it
BRAG’s Planner emits a DecompositionPlan — a list of evidence slots, each tagged with an evidence_type like strategic_position, forward_guidance, operational_metric, or accounting_policy. The retriever filters candidates by type before scoring. The Verifier grades coverage against the right rubric for that type.
Apply it in your own system
You don’t need BRAG’s specific taxonomy. The move is just: stop treating all chunks as equivalent.
- Build a taxonomy of evidence types relevant to your domain (5–10 categories is usually enough).
- Tag documents and chunks at ingest time with one or more types.
- Have your planner decide what kind of evidence each question needs.
- Filter retrieval candidates by type before similarity scoring.
This is one of six principles abstracted from BRAG — a retrieval‑augmented generation system with two LLM skeptics gating every confident answer. Tested on ten years of Netflix’s public financial reporting.