Retrieval systems fail quietly
A retrieval system that returns the wrong passage produces a confident, well-written, wrong answer. Nothing in the output signals the failure, which is what makes it the dangerous kind.
Thomas Reinhardt, Head of AI Services ·
The failure has no signature
When generation fails, the output is visibly wrong. When retrieval fails, the output is fluent and sourced from the wrong place. Users cannot tell these apart, and neither can a generic benchmark.
Cite the passage, always
Rendering the source paragraph alongside every answer converts an invisible failure into a visible one. It costs screen space and saves the system.
The corollary: if there is no supporting passage, return nothing. An honest empty result is more useful than a plausible invention, and it is the behaviour most often designed out under pressure to look capable.
Tune against your own questions
Chunking strategy, embedding choice and reranking all interact with the shape of your documents. Benchmarks measured on someone else’s corpus transfer poorly. Build the question set from the real archive and tune against that.