AI features

How AI Query Understanding and AI Search Coach work in practice.

Skryx ships two AI features that improve search without you having to do anything:

  1. AI Query Understanding — rewrites vague queries on the fly.
  2. AI Search Coach — analyses your search analytics every day and surfaces actionable recommendations.

Both are part of the Growth plan and above. Neither requires extra code.

# AI Query Understanding

When a customer types something keyword-search can't handle (headphones for travel, ceva pentru tractorul meu care nu porneste, laptop for video editing), Skryx asks the AI to rewrite the query into precise search terms and detect intent. The rewrite is used internally; the original query and a user-facing message come back in the ai_context field of the response.

"ai_context": {
  "original_query": "headphones for travel",
  "rewritten_query": "noise cancelling wireless headphones",
  "intent": "buy_part",
  "user_message": "Looking for portable, noise-cancelling headphones for travel",
  "alternative_queries": ["portable bluetooth headphones", "travel headphones"],
  "confidence": "high"
}

# When it fires

A built-in heuristic decides whether to call the AI:

  • Skip for SKUs, single-word queries, specific brand+model strings (cost-optimisation; ~70–80% of queries).
  • Use AI for queries with use-case markers (for , best, how, ?) or queries longer than 3 words.

You can tune this per tenant in Settings → AI Query Understanding:

  • complex_only — the default heuristic.
  • all — every non-trivial query goes through AI (more cost, broader coverage).
  • off — disable entirely.

# Caching

The first time a query is rewritten, the result is cached for 7 days. Subsequent identical queries — including across different visitors — are served from cache in < 5 ms with no AI cost. Expect 60%+ cache hit rate after a few days of real traffic.

# Costs

A single rewrite costs roughly $0.0005–$0.0015 in AI usage. With caching, the effective per-query cost in production is well under a cent per 1,000 queries. The monthly cap in the table on the rate limits page is generous compared to typical traffic.

# UI integration

In the search response, when ai_context is non-null:

  • Show user_message as a small banner: "Looking for portable, noise-cancelling headphones for travel."
  • Offer a Search original link if you want to let users opt out.
  • Render alternative_queries as one-click chips below the banner.

The Skryx Playground does exactly this — copy the pattern from there.

# AI Search Coach

Coach runs an analysis on your search events and produces structured recommendations:

  • Missing synonyms — terms customers searched but didn't match anything (e.g. placute frana → pastile frana).
  • Catalog gaps — terms with meaningful volume and zero relevant products.
  • Ranking opportunities — popular queries where the top hits are likely a poor fit.
  • Top performers — your best-converting queries.

Each recommendation has a one-click Apply that does the actual work: creates the Synonym row, pushes the synonym set to the engine, or creates the Ranking Rule. Dismiss keeps the recommendation out of suggestions for 30 days.

# Scheduling

By default, Coach runs:

  • Daily at 04:00 UTC for a 24-hour pulse.
  • Weekly Monday at 05:00 UTC for a deeper 7-day pass.

You can also run it on demand from the dashboard or via POST /v1/indexes/{name}/coach/analyze.

# Costs

A 7-day analysis on ~250 search events costs around $0.02 in AI usage. The monthly quota is generous; you'll usually use a fraction of it.

esc