Two searches run in parallel over 758 manual sections: BM25 (exact words — error codes, menu labels) and embedding similarity (meaning — finds sections even when words differ).
Reciprocal Rank Fusion merges both rankings; a section ranked high by both wins.
Confidence gate: if no section is semantically close enough, the API answers confident: false — the voice agent then says "I don't know" instead of guessing. Try the off-topic chip.
Deep links: anchors are extracted from the published manual at ingestion, so every link lands on the exact heading at help.softmedic.sk.
Same API the ElevenLabs voice agent calls as a tool; every query is audit-logged with retrieved chunk IDs and manual version.