Skip to course content
Free RAG fundamentals course

Introduction to RAG and Knowledge Assistants

Unit 03.01: Why similar is not the same as correct

The previous unit showed similarity finding a paraphrase. This one shows the same mechanism finding a contradiction and treating it as a match.

Two chunks, opposite claims, near-identical scores

Similarity ranks by subject matter. Two sentences stating opposite policies are about the same subject, use the same vocabulary, and therefore sit almost on top of each other in the vector space.

The table below scores both against the same query. Look at how close the two numbers are before reading further.

c1 similarity 1.000  Refunds are allowed within 7 days.
c2 similarity 1.000  Refunds are NOT allowed on sale items.

Similarity ranks by topic. Contradiction is invisible to it.

Both score near 1.0. There is no signal in the geometry that they disagree - negation, contradiction and mutual exclusivity are not properties an embedding was trained to represent.

So the retriever hands the model both, and the model picks one, and nothing in the pipeline records that a contradiction was present. That is not a tuning problem; a better embedding model makes the two chunks score *closer*, not further apart. Detecting the conflict requires a separate check, which Module 7 builds.

The mistake this prevents

The mistake is expecting a retrieval quality metric to catch this. Recall@k is satisfied - the correct chunk was retrieved. It was retrieved alongside its own contradiction, and the metric has no opinion about that. Conflicts have to be looked for explicitly.

Takeaway

Similarity finds subject matter. It has no notion of agreement, negation or truth, so a retriever will return contradictory chunks with equal confidence and leave the resolution to something downstream.