A RAG system confidently quotes a return-policy clause that was deleted last month. The most likely cause is:

Stale chunks in the vector store that no one reindexed.

Embeddings have too few dimensions.

Cosine similarity is the wrong metric.

When RAG fails — Semantic Web Academy

Overview

Hallucination, stale chunks, lost context.

Why it matters

Knowing the failure modes by name is half the cure. Each maps to a concrete fix covered in later levels.

Where this sits in the stack

When RAG fails is one of the load-bearing decisions in a KG/RAG/agent system: choices made here propagate to retrieval quality, agent reliability, cost per query, and the on-call burden of whoever ships it. Teams that name this trade-off explicitly ship faster than teams that leave it implicit.

Why this is load-bearing

When RAG fails is the building-code of this layer. You can ignore building codes on a shed, but the moment you put two storeys on top of the same foundation they decide whether the structure stands or falls. In a KG/RAG/agent stack, the equivalent of 'two storeys' is the second feature you ship on top of this primitive — GraphRAG on top of chunking, supervisor agents on top of state machines, regression CI on top of metrics. The cost of cutting the wrong corner now is paid by every later layer, with interest.

Knowing the failure modes by name is half the cure. Each maps to a concrete fix covered in later levels.

Reflect — apply it

Anchor when rag fails to something concrete in your own work.

›Where have you seen when rag fails done well? Name one team or product and what they got right.
›Where have you seen it done badly? What was the first symptom that surfaced (latency, hallucination, cost, outage)?
›What is the *cheapest* version of this you could ship in your next sprint, and what single metric would tell you it's working?

Reading in progress · 0 of 3 activities done