SearchSaver

An AI agent that finds your store's zero-result searches, auto-fixes them with Elasticsearch synonym & spelling rules, and shows the revenue you recovered.

Comment

Inspiration

Every e-commerce site silently loses money to failed searches. A shopper types "trainers," the catalogue says "running shoes," and they get zero results — so they leave. These misses are invisible: they don't show up in a sales report, nobody files a bug, and most stores never measure them. We wanted an agent that surfaces this hidden leak and closes it — not with a dashboard a human has to act on, but by generating and applying the actual Elasticsearch fix.

What it does

SearchSaver connects to any Elasticsearch cluster and:

  1. Finds every search from the last 30 days that returned zero results, ranked by frequency.
  2. Classifies each one with Gemini 2.5 Flash: SYNONYM (trainers → running shoes), SPELLING (adibas → adidas), CASE (NIKE AIR), or MISSING (genuinely not in the catalogue — unfixable).
  3. Generates real Elasticsearch analyzer settings — a synonym_graph filter plus pattern_replace filters — and on your own cluster applies and verifies them (re-indexes a .search sub-field and confirms the query now returns results).
  4. Quantifies the impact: it estimates and animates the monthly revenue recovered per fix (query frequency × conversion rate × average order value).

The hosted demo runs live against a real Elastic Cloud cluster seeded with 48 products and 608 search events.

How we built it

  • Google ADK 2.1.0 orchestrates the agent; the model is Gemini 2.5 Flash via Vertex AI.
  • Elastic is the partner integration: Elasticsearch 8.13 holds the catalogue + search logs, and the Elastic MCP server runs as a second Cloud Run service for cluster inspection.
  • A FastAPI backend streams progress to a vanilla-JS frontend over Server-Sent Events.
  • Deployed on Google Cloud Run (two services: the app + the Elastic MCP server).
  • Key decision: the agent is tool-free. A Python pre-fetch layer runs the server-side aggregations, hands the model a compact brief, and lets it only reason and emit JSON. Writes go through a dedicated apply step over httpx.

Challenges we ran into

  • Malformed tool calls. Giving the agent the Elastic MCP toolset directly made Gemini attempt — and ~1/3 of the time malform — tool calls it didn't need, the main cause of failed analyses. Switching to the pre-fetch pattern took reliability from ~2/3 to 5/5 and cut end-to-end time from ~90s to ~25s.
  • Elasticsearch won't let you change a live field's analyzer, and synonym_graph can't run at index time. The fix builds a separate search analyzer and an analyzed .search sub-field, then runs _update_by_query to populate it — so multi-word synonyms actually match.
  • Free-tier model limits capped the AI Studio key at ~20 requests/day and broke the live demo, so we moved to Vertex AI with a service account.
  • A safe public demo: visitors can analyse the shared cluster freely, but writes are locked and the API key is injected server-side, never exposed.

What we learned

For agentic reliability, don't give the model a tool it doesn't need. Pre-computing the data deterministically in code and reserving the LLM for the reasoning step it's uniquely good at was faster, cheaper, and far more reliable than a fully tool-driven loop.

What's next

  • Scheduled re-scans with email alerts on new zero-result spikes.
  • Auto-tuning conversion-rate assumptions from each store's real CTR.
  • One-click rollback of applied analyzer settings.

Built With

Share this project:

Updates

Submission history