Use case · Operations & Edge

Search that understands “waterproof jacket under $100”.

A small model fine-tuned on your catalog turns natural-language queries into structured filters and ranked results. Self-hosted next to your existing search engine, answering in under 100 milliseconds — at a flat monthly cost, not a per-query API fee.

Natural-language catalog search puts a small model fine-tuned on your catalog in front of your existing search engine. It translates free-text intent into structured filters and a cleaned query in tens of milliseconds, self-hosted in your VPC, while Elasticsearch, OpenSearch, or Algolia keeps doing retrieval and ranking. No re-platforming, no per-query API fee — and we measure the result against your own search logs, before and after.

Shoppers type intent. Keyword search returns the wrong aisle.

“Warm mid-layer for hiking, not down.” “Office chair for a bad back under $300.” Keyword search returns zero results or the wrong category. Frontier APIs parse that intent, but at per-query prices and 1–2 second round trips that don't belong in a search box — so most catalogs either stay on keywords and lose the sale, or pay frontier rates for what is, per query, a narrow translation task.

How it works · The Crit Path
01Scope — start from your search logsIn The Crit we pull zero-result queries, reformulation chains, and abandonment points from your existing analytics. That corpus defines both the training data and the frozen eval set of real queries.
02Prove — fine-tune a 1–3B model on your catalogThe model learns your taxonomy, attributes, and price structure, and translates a free-text query into structured filter JSON — category, attributes, price bounds — plus a cleaned query for retrieval. Constrained decoding keeps the output valid against your schema.
03Deploy — next to your engine, not instead of itThe model serves from a single GPU in your VPC via vLLM; Elasticsearch, OpenSearch, or Algolia stays the retrieval and ranking engine. The small model is a translation layer in front — your search infrastructure doesn't change.
04Assure — measure against the log baselineZero-result rate, click-through on the first results page, and p95 latency, compared to the pre-launch numbers from your own logs. If the model doesn't move them, you see that on the same dashboard we do.
The economics

A search box has a latency budget — and a per-query fee breaks the cost one. Illustrative math from our cost model and public list prices, August 2026; The Crit replaces it with your query volumes: at 1M searches a month, parsing intent through a frontier API at $0.005–0.015 per query costs $5,000–15,000 a month, with 800 ms+ round trips that force aggressive caching or a degraded experience. A fine-tuned 1–3B model on a single L4-class GPU runs $400–700 a month flat — roughly 10–20× less at this volume, by the same illustrative math — and answers in tens of milliseconds. The pattern matches published results elsewhere: Forethought reports 66–80% lower inference costs after moving routine workloads to fine-tuned small models (on AWS), and Gartner expects task-specific small models to see three times the adoption of general-purpose LLMs by 2027.

Frontier API per queryFine-tuned model in your VPC
Cost shapePer query, linear with trafficFlat: one mid-range GPU, regardless of query count
Latency800 ms+ round trips — outside a search box's budgetTens of milliseconds — inside the budget
Data exposureEvery customer query streams to a third partyQueries never leave your VPC
Your stackCache layers and workarounds bolted onElasticsearch, OpenSearch, or Algolia stays — the model is a layer in front
Where this doesn't work

If your catalog runs under a few thousand SKUs with low search traffic, better keyword synonyms probably solve it — and The Crit will say so. We also make no revenue-lift promises: conversion effects depend on your catalog and traffic, so we show you zero-result and click-through deltas on your own dashboard instead of quoting someone else's.

What you get

A layer in front. Not a re-platform.

  • Own the query-understanding model trained on your catalog and search logs — the weights are yours
  • Deploy it in your VPC as a layer in front of your existing search engine
  • Keep a frozen eval set of real customer queries, plus a written accuracy report
  • Measure zero-result rate, first-page click-through, and p95 latency against your log baseline
  • Retrain from a runbook when the catalog or seasonal assortment changes
Who this is for
Retailers with real search volumeRoughly 100K+ queries a month and a visible zero-result or search-abandonment problem in the analytics.
Teams that keep their engineElasticsearch, OpenSearch, or Algolia already in production — intent understanding added in front, not instead.
Queries that stay homeCustomer search terms that never leave your VPC on the way to a result.
Questions, answered straight

Do we have to replace Elasticsearch or Algolia to add AI search?

No. The small model sits in front of your existing engine as a translation layer: it converts a natural-language query into structured filters and a cleaned search string, and your current engine does the retrieval and ranking as before. No re-platforming, no index migration.

What does self-hosted AI product search cost to run?

A fine-tuned 1–3B model handles production search traffic from a single mid-range GPU — roughly $400–700 a month at current cloud GPU prices, flat, regardless of query count. The alternative — parsing each query through a frontier API — scales linearly with traffic: at 1M queries a month, roughly $5,000–15,000 by our illustrative math from public list prices, August 2026. The Crit replaces these assumptions with your actual volumes.

How do you measure whether AI search actually works?

Against your own logs, before and after. We baseline zero-result rate, first-page click-through, and reformulation rate from your existing search analytics, then track the same metrics after launch — plus model accuracy against a frozen eval set of real customer queries, which we gate before rollout. No industry-average claims; your dashboard is the evidence.

Security · by architecture

Customer data stays home.

  • Conversations never leave
  • Card data can’t reach a prompt
  • No new subprocessor
  • Region-pinned by design
Architected to deploy inside your
PCI DSS GDPR CCPA
Small models · critical hits

Your search logs already list the lost sales.

Zero-result queries, reformulation chains, abandonment points — The Crit turns them into a baseline, a fine-tuned model, and a dashboard that proves the difference on your own numbers.