Experiment: hybrid (semantic + keyword) retrieval for the livesearch suggestions

[reekitconcept]

Agreed on the intranet ticket 570 review call (2026-09-08): worth trying as a separate experiment, decided on results — not part of the current alpha stream.

Question

@solr-suggest is pure keyword/ngram today (suggest, suggest_ngram, searchwords fields). @rag-search already does hybrid retrieval (knn vector + BM25, RRF-fused). Should the suggestions also use the hybrid retrieval, so that all three searches (results / suggestions / AI) rank consistently and semantic matches surface while typing?

Why it's cheap to try

All the pieces exist in rag/pipeline.py: embed_query, search_chunks (permission-trimmed, honors path_prefix and extra_conditions), parent_ranking, rrf_fuse. Roughly 1–2 days plus tuning.

Known concerns (to be answered by the experiment)

• Hot path: suggest fires per (debounced) keystroke; each hybrid call needs an embedding round-trip (latency; and each LLM call occupies a backend worker thread — at keystroke frequency, far more load than per-question RAG).
• Partial words embed uselessly: typing vaca should complete to vacation (the ngram fields' job), not find semantically-near content of an unfinished word.
• Explainability: semantic hits can share no text with the query — highlighting breaks, users ask "why is this suggested?".
• Availability: suggest must keep working with the LLM stack down → silent keyword-only fallback required.

Experiment shape (proposal)

• knn leg only for complete-looking queries (multi-word, or trailing space); prefixes stay ngram-only.
• Debounce ≥300 ms; hard embedding budget (~150 ms) with silent keyword-only fallback; per-query-string embedding cache.
• UI idea (David): keyword suggestions by default, plus an explicit "semantic" toggle/button once the full query is typed.
• Evaluate on the German intranet corpus: latency, load on the LLM endpoint, and whether the semantic additions match what users would expect — the deciding criterion is user expectation, not technical elegance.

Outcome

A recommendation with examples (including negative ones if it works worse), then the team decides whether to productize.