Skip to content
Quantoros

Why does standard similarity-based RAG fail on 'how do X and Y relate across three documents?'

LLM & AILevel: IntermediateRAG (Retrieval-Augmented Generation)

Why does standard similarity-based RAG fail on 'how do X and Y relate across three documents?'

Hints
  • Break the query into candidate retrieval, context selection and generation.
  • How would you locate the stage that lost a fact without using only one end-to-end metric?

Answer

The failure risk is evidence coverage. A similarity query may retrieve passages separately describing X and Y while missing a bridge entity, causal link or intermediate fact needed to connect them. Independent chunk scores do not guarantee that the returned set contains all premises of a multi-hop answer.

Failure is not inevitable: if the needed facts are retrieved, a generator can synthesize across documents. Possible improvements include decomposing the question into subqueries, iterative retrieval, retrieving linked entities or neighboring passages, and graph-based indexes. Evaluate whether the evidence set supports every reasoning step. GraphRAG is one approach for connected or corpus-level questions, not a universal replacement for simple retrieval.

Source and licence: Skillumen — Awesome AI Engineer Interview Questions · CC0-1.0 · answer: Quantoros editorial

Related questions