What does it need to work?

A view of the company’s entities and of the asker’s own context — who they are and what they work on — so pronouns, shorthand and implied timeframes resolve to something specific.

What does it look like when it is missing?

Plausible answers about the wrong entity or the wrong period, delivered with exactly the same confidence as the right ones.

Query Understanding

Query understanding is working out what a question actually asks — which entity, which timeframe, what intent — before anything tries to retrieve an answer for it.

Questions people ask at work are short and lean heavily on context. "How did the Acme renewal go?" carries an entity that has to resolve to a record, an implied timeframe, and an intent that could be about the outcome, the conversation or the numbers. Retrieval against the raw sentence treats all of that as keywords.

The step matters most when it changes the retrieval rather than merely re-ranking it. Resolving "Acme" to an account means fetching that account’s records instead of everything mentioning the word. Recognising a comparison means running two retrievals instead of one. Systems that skip this stage tend to be good at questions that are already well-formed and poor at the ones people actually type.

Frequently Asked Questions

It is the step before retrieval in a well-built RAG system, and the step most often left out. Embedding the raw question and searching is the simplest version, not the good one.