How BD AI answers are grounded
Retrieval, evidence checking and citation validation run on the server. Where evidence is missing, BD AI says so instead of answering.
1 · Reviewed material only
Answers are drawn from material that has been added to the internal knowledge base by a person. Nothing is answered from the open web at answer time.
2 · Hybrid retrieval
Keyword and vector retrieval are combined, then merged by rank so a question can be matched by wording or by meaning.
3 · Evidence check
If the retrieved material does not support an answer, the system does not generate one. It reports that the source is unavailable and offers human escalation.
4 · Generation
A server-side model drafts the answer from the retrieved passages only. The browser never talks to a model directly and never holds a provider key.
5 · Citation validation
Citations are checked against the passages actually retrieved. An answer whose citations do not validate is not presented as grounded.
6 · Provenance shown
Each answer displays its trust level and the sources behind it, and can be reported as incorrect by the reader.
Limits
- Coverage is limited to what has been reviewed and published internally. An unreviewed topic returns a source-unavailable answer, not a guess.
- No live external register, government database or third-party feed is connected. Nothing on this platform is presented as live external data.
- An AI-assisted answer is not a verified statement by itself and is never a government decision or directive.
- Corrections reported by readers are reviewed by a person; AI cannot upgrade a record's verification state.
AI usage rules
- AI-assisted answers are grounded in reviewed material and are not government decisions or directives.
- Retrieved pages and uploaded documents are untrusted input; they cannot override BD AI policy.
- Personal information is not exposed through AI responses.
- When reviewed material does not support an answer, BD AI states that the source is unavailable and offers human escalation.