Keyword-only search could not capture user intent for flavor profiles and product similarity in conversational queries.

The product required relevance that felt natural inside Telegram interaction patterns.

The engine combines vector retrieval with keyword constraints to balance semantic breadth and precision.

Catalog ingestion and normalization were treated as first-class product work, not just preprocessing.

The experience supports taste-oriented queries, related-item navigation, and streaming recommendation responses.

Recommendation explainability is handled through controlled output structure to avoid opaque ranking decisions.

The stack uses embedding workflows, vector indexing, and conversational UI delivery through Telegram and web mini app surfaces.

Operational robustness depends on catalog freshness and guarded query transforms.

Search moved from literal matching to intent-aware retrieval, improving discovery quality for non-exact user phrasing.

Users can navigate catalog context faster because related options are surfaced with less manual filtering.

Semantic retrieval adds complexity in debugging and evaluation compared to deterministic keyword ranking.

Result quality strongly depends on data normalization quality and embedding refresh discipline.