Online shops whose search bar only matches keywords.
What it costs today
Shoppers type what they mean ("a waterproof jacket for hiking under $200") and get nothing, or the wrong page, and leave.
What solved looks like
Search that understands the request, still finds exact model numbers and SKUs, and answers with real products and links.
Meaning and exact codes in one search: vector and keyword together, reranked.
Every suggestion linked to the product page it came from.
Optional live stock and price through your store's API.
How it runs
Product search in plain language step by step
Each stage is a real step in SentientOne. Timings are medians from our latest benchmark; your model and network change them.
Product search flow. Your shop's search box sends the shopper's words to the REST API from your server. 1: your secret key and the agent are checked and one credit is reserved. 2: popular first questions are answered from the answer cache in about 46 milliseconds. 3: your crawled catalogue pages are searched by meaning and by keyword, fused and reranked. 4: the agent can check live stock on your store's MCP server. 5: a fast model writes a short answer with the matching products. 6: the answer and its sources stream back, and your page shows the sources as product links.
Channel
Your search box, through the REST API from your server
Shopper query
“Waterproof hiking jacket under $200, size M”
Context
conversation_id for follow-ups
Agent: Product Discovery
X-Request-Id
SENTIENTONE
Auth, credits
Preflight
Your secret sk-so- key from your server; the agent's limit; 1 credit reserved.
60 req/min · 402 at zero
Answer cache
~46 ms
Popular first questions to an unchanged agent come straight from cache.
15-minute TTL · cleared on re-crawl
Knowledge
~450 ms
Crawled product pages: meaning and exact SKUs in one hybrid query, reranked.
shop.example/products · 200 pages per crawl
MCP: stock
Optional
Optional: the model checks stock and price on your store's MCP server.
check_stock · 10 s per call
Generation
Short answer
A fast model picks the matches and says why each fits.
GPT-5.4 mini on your key
Sources
SSE
The sources event lists each product page used, so you render them as cards.
title · url · score · snippet
Search answer
Done · 1 credit
Two jackets fit: the Trailhead Shell (SKU TH-220, $179), fully waterproof and light [1], and the Ridge Lite ($149), better for drizzle than downpours [2]. Both come in M.
1Trailhead Shell · shop.example/products/th-220
2Ridge Lite · shop.example/products/ridge-lite
trace · 2 sources · 1 credit · x-request-id from your server
Channel
Your search box, through the REST API from your server
Shopper query
“Waterproof hiking jacket under $200, size M”
Context
conversation_id for follow-ups
Agent: Product Discovery
X-Request-Id
query
SENTIENTONE
Auth, credits
Preflight
Your secret sk-so- key from your server; the agent's limit; 1 credit reserved.
60 req/min · 402 at zero
credit
Answer cache
~46 ms
Popular first questions to an unchanged agent come straight from cache.
15-minute TTL · cleared on re-crawl
miss
Knowledge
~450 ms
Crawled product pages: meaning and exact SKUs in one hybrid query, reranked.
shop.example/products · 200 pages per crawl
top 5 chunks
MCP: stock
Optional
Optional: the model checks stock and price on your store's MCP server.
check_stock · 10 s per call
in stock
Generation
Short answer
A fast model picks the matches and says why each fits.
GPT-5.4 mini on your key
tokens
Sources
SSE
The sources event lists each product page used, so you render them as cards.
title · url · score · snippet
cache hit · first token ~46 ms
first token ~1.1 s with a knowledge search
sources + text
Search answer
Done · 1 credit
Two jackets fit: the Trailhead Shell (SKU TH-220, $179), fully waterproof and light [1], and the Ridge Lite ($149), better for drizzle than downpours [2]. Both come in M.
1Trailhead Shell · shop.example/products/th-220
2Ridge Lite · shop.example/products/ridge-lite
trace · 2 sources · 1 credit · x-request-id from your server
Set it up
The walkthrough, in 8 steps
1
Add your LLM key
Agents run on your own model provider key; there's no shared one, so nothing answers until you add it. Create a key with OpenAI, Anthropic, Google Gemini or Groq, then open Configure → LLM keys → Add key. It's encrypted at rest with AES-256-GCM, used only to call your provider, and never shown again in full. If a reply fails with INTERNAL_ERROR straight away, the agent has no key; with LLM_AUTH_ERROR, the provider rejected it.
Agents → New agent → Start from: Product Discovery. Pick a fast model for search (for example GPT-5.4 mini or Gemini 2.5 Flash) and pick your saved LLM key.
Tell it to recommend only products from the knowledge base, to keep answers short, and to say when nothing fits.
System prompt
You help shoppers find products in our catalogue.Recommend only products that appear in the knowledge base. Give the name, the SKUand the price, and one line on why it fits. At most three products.If nothing fits, say so and suggest the closest category instead.Never invent stock, sizes or prices.
Knowledge → Websites: add your shop's address with an include path such as /products and max pages set to 60 (300 credits). Each crawl stops at 200 pages, so add one crawl per category if you have more. Re-crawl when prices or ranges change; the old pages are replaced and charged again.
If your store has an MCP server for stock and prices, add it on the Tools tab and Test connection. Note that agents with live tools skip the answer cache, so repeated searches won't be instant.
Open Tracing to see every request: what knowledge it found, each tool call, tokens and timing, and the error code if it failed. Analytics charts requests, errors and response time per agent. Under Billing, turn on auto top-up so the agent never stops for want of credits.
Tell us about your setup and we'll help you build product search in plain language on SentientOne: the agent, the knowledge, the tools and the deployment.