Semantic retrieval
Answers you can trace back to the sentence.
Every answer is grounded in your corpus, scored for how well the sources support it, and linked passage by passage to where it came from.
How an answer is built
Grounding is a sequence of checks.
Each stage is a setting you own, and each can refuse. An answer that cannot clear them routes to a fallback or declines.
Hybrid search over your corpus
Keyword, vector or hybrid retrieval against your own knowledge base, with the query type chosen per deployment. Custom vector queries are supported where your search backend offers them.
Source selection under a threshold
A document score range decides which sources are eligible, a date penalty deprioritises stale ones, and a cap on documents per call bounds what reaches the model.
Semantic re-ranking and coverage
Retrieved passages are re-ordered by how well each one matches the answer given. Answers stitched across many documents are penalised, and the minimum coverage for the top source is yours to set.
Sentence-level grounding checks
Proper nouns and other search terms in the answer are checked back against the sources. Sentences containing key words the knowledge base does not contain can be removed before delivery.
Confidence gates you set
A warning threshold prepends a caveat; a minimum threshold substitutes your reply text or hands off to a fallback agent; a separate floor decides whether a link is shown at all.
Provenance as part of the answer
Provenance highlighting shows which parts of an answer came from which source, sources are attributed by document name, and a link field returns the URL — above the confidence floor you set for showing one.
Grounded and cited
Measured, not asserted.
Each answer returns a semantic match score, the knowledge base's own confidence and coverage, and the response-time breakdown behind it. The highlighted phrases are the passages the answer was built from — on screen, not in a report.
Refuse below a floor
What a decline looks like.
Below the minimum confidence you set, the platform substitutes your reply text or hands the question to a fallback agent instead of guessing — and the log records the score, the intent and the reason.
Route it onward
The cache
The same question, asked many ways.
Questions resolve to intents, and an intent's curated answer can be served without a model call. Every intent shows its coverage and confidence over time, flags when its sources changed, and can be edited by a subject-matter expert.
Matched to an intent
Three questions this page exists to answer.
They are the ones that come up in the second meeting, once the demo has already worked.
How do we know an answer is actually grounded?
Because grounding is measured. Each answer carries a semantic match score, the knowledge base's confidence and coverage, and the passages themselves. Those are thresholds: an answer under yours is substituted or routed away, and Governance keeps the distributions over time.
What happens when the corpus does not contain the answer?
It says so. A minimum-confidence floor turns a weak answer into your reply text, and the fallback can be a template that notifies a team or opens an issue. The decline is logged either way.
Does caching mean we serve stale answers?
Only if nobody is watching. Cached answers carry a duration you set, intents are flagged when their source documentation changes, and the Curate view filters for edited, flagged, new and out-of-date entries. Curation is an editorial task with a record.
See it running inside your own boundary.
Talk to a NeuralSeek expert about how it fits your stack, where your data has to live, and the governance your auditors already expect.