Production RAG

It answered well in testing.
A month later it degraded.

Vector similarity is the first stage of a retrieval strategy, not the whole one. We build the stages that come after it.

Hybrid searchRe rankingGuardrailsDrift detection

 retrieval quality by pipeline stage ●measured 
 vector search alone recall to partial 
 plus hybrid and rewriting recall to better 
 plus cross encoder re rank recall to 99%+ 
Sub 100ms across more than a million documents in production, with chunking that respects semantic boundaries and guardrails before anything reaches a user.

How a working RAG system goes wrong.

Confidently wrong answers

Retrieval returns something plausible from the wrong document and the model presents it without hesitation.

Quality that decays quietly

Embeddings drift as content changes. Nobody notices for six weeks because nothing measures relevance over time.

Latency nobody budgeted

Each added stage costs milliseconds. Without a latency budget the pipeline creeps past the point users tolerate.

Escalations climbing

Support volume goes up rather than down, because the assistant answers just well enough to waste the customer time.

What we actually do here.

Multi stage retrieval, cross encoder re ranking, query rewriting, content guardrails and the monitoring that catches drift before your users do.

See the whole stack

Measure recall honestly

A labelled evaluation set from your own content. Without it, every change to the pipeline is a guess.

Fix chunking and hybrid search

Chunks that respect semantic boundaries, keyword and vector blended, so exact matches stop getting lost.

Re rank and guard

A cross encoder on the shortlist, query rewriting for vague inputs, and safety checks before the response ships.

Watch for drift

Embedding monitoring, relevance feedback loops and alerts that fire on quality, not just uptime.

Find out whether your retrieval is degrading.

The scan covers your distribution footprint in 40 seconds. For a retrieval audit, the founders read your pipeline by hand.

P.S. If the assistant is answering confidently and wrong, that is the expensive kind of quiet.

It answered well in testing.A month later it degraded.