Answered by default#
Type your query and the portal reads the top results and writes a short, cited answer over them, with the sources it drew on listed underneath. Every claim carries a citation you can follow back to the passage it came from.
The ranked documents are always there below the answer, each with the passage that matched, so you can judge a source and open it yourself.
Results only turns the answer off when you just want the document list - it is instant and answer-free. Press it again to bring the answer back. From an answer you can Continue in Ask to keep asking follow-ups.
Retrieved and cited - the difference#
An answer is built in two steps, and the portal shows you both:
- Retrieved is what search found for your question. It is the whole pool the answer was written from, including passages that turned out not to be useful.
- Cited is the smaller set the answer actually drew on - the numbered [1] markers in the text. Each one links to the exact passage it came from.
So every cited source was retrieved, but not every retrieved source is cited. Seeing both is deliberate: the cited set tells you what the answer rests on, and the wider retrieved set lets you check whether anything relevant was found but passed over.
How matching works#
Three retrieval modes sit under the box:
- Hybrid (the default) combines keyword and semantic matching - the best all-round choice.
- Semantic matches on meaning, for when the right words are hard to pin down.
- Keyword matches on the exact terms, for a known phrase, code or name.
Open Filters to narrow by topic or by document kind (for example Report or Submission). The Match strength control switches between All results and Strong ones only - a strong match scores 60% or higher on the calibrated relevance scale.
Reading results honestly#
Each result carries a relevance score on a calibrated 0 to 100 scale, so a weak match looks weak rather than being inflated to the top. Results below the noise floor are dropped, and a query the corpus cannot answer honestly returns nothing rather than surfacing irrelevant hits.
Reference lists and bibliographies stay findable but never outrank real body text, and near-duplicate pages are collapsed so you do not see the same content twice.
More you can do#
- Summarise these results writes a quick synthesis across the current result set.
- Watch this search keeps the query and re-checks it daily - a dot appears next to the saved search when new results turn up.
- People also ask suggests related questions when the corpus has ones that genuinely overlap with your search.
- Save on any result adds it as evidence to your current investigation.