No content blocks yet.
Retrieval Is Not Intelligence
Your RAG system plateaued because similarity search answers lookup questions, and the questions worth money are assembly questions. A failure taxonomy.
Key Takeaways
- RAG systems plateau in the low eighties because the residual failures are mostly assembly questions - aggregation, joins, temporal diffs, absence - whose answers exist in no retrievable chunk.
- Retrieval is a resemblance detector, and assembly questions have a question-evidence gap: their evidence doesn't resemble them, so retrieval fails quietly by returning question-shaped material instead of answer-bearing material, which the model fluently synthesizes into misleading partial answers.
- The question ladder (lookup → multi-hop → aggregation → temporal → absence → synthesis) locates each failure and the capability it demands.
- Breaking the plateau means answering at write time: extract structure at ingestion - knowledge graph edges for joins, derived tables for aggregates, versioned states for diffs - and demote retrieval to an evidence-citation tool inside an agentic loop for the long tail.
- Absence deserves special care: it's a closed-world claim requiring enumeration and explicit epistemics, not a retrieval result.