Semantic search for large catalogs
A hybrid BM25 + embeddings engine for stores with thousands of SKUs. It copes with typos, respects exact part codes and records which searches end in an order. It plugs into Shopify, PrestaShop or a custom store through a widget.
76,000+
Indexed products
~60 ms
Per search
BM25 + AI
Hybrid engine
For stores that have outgrown their built-in search
The search that ships with your platform matches letters. Give it a typo, a plural or a phrase like "something to help me sleep" and it returns nothing, or products that have nothing to do with the query. AgentSearch indexes your catalog in Typesense, ranks with BM25 and brings in embeddings where they help. We deploy it and look after it, and the search data stays in your own database: what people look for, what they can't find and what ends up in an order.
Relevant results from the very first keystroke
Suggestions as you type, typo tolerance and a ranking that takes into account which field each word matched.
- Hybrid BM25 + vector search for queries that share no words with the product page
- Autocomplete: the last word is treated as a prefix while typing
- Up to two typos per word, plus editable synonyms
- Responses in ~60 ms even with catalogs of 76,000+ products
A search chat that recommends products
The AI layer (Azure OpenAI) is optional and switched on per store. It understands natural-language queries and chats to recommend products from your catalog.
- Reads natural-language queries and works out the intent
- Finds products, checks stock and adds to cart from the chat
- Suggests alternatives when a search comes back empty
- Won't claim a part fits a model unless the product data confirms it
Installed with a snippet in your theme
A JS widget (Shadow DOM, about 23 KB gzipped, no React) that hooks into the search box of Shopify, PrestaShop or a custom store, with facets that adapt to your catalog.
- Its own Shadow DOM: your theme's CSS can't touch it
- Facets by brand, category, type and attributes
- Curations to pin or hide products for a given query
- Search by photo and a similar-products button
What the engine includes
One engine for relevance, AI, store integration, merchandising control and measurement.
Hybrid search
Typesense with BM25 for keyword relevance and 3,072-dimension embeddings for meaning. BM25 leads on brands and part codes; vectors come in where shoppers type full sentences.
AI layer and chat
Optional and switched on per store: it suggests alternatives when there are no results, interprets vague queries and runs a chat that finds products and checks stock. Spending has a daily cap.
Embeddable widget
A script with no React that renders inside its own Shadow DOM and hooks into your theme's search box. About 23 KB gzipped, with several languages built in.
Facets, synonyms and curations
Filters by brand, category or attribute that recount the rest when you tick one. One-way or equivalent synonyms, and curations to pin or hide products for a query, with start and end dates for campaigns.
Search analytics
Queries, zero-result searches, clicks with their position, add-to-carts and attributed orders with their value. Stored in your database, with a daily report of what shoppers couldn't find.
Several stores, one index
A single index serves several stores and the internal back office, and each one sees its own stock, price and URL. That's how it runs in production today across several Shopify stores sharing one catalog.
From your catalog to live search in three steps
No store rebuild. The slow part is usually tidying up the catalog data.
We index your catalog
We read products from the store, the PIM or the ERP with read-only access and load them into Typesense with their vectors. From then on, an incremental job keeps them up to date every few minutes.
We tune relevance and AI
We adjust field weights, facets and synonyms against real queries, and decide which stores get the AI layer. The key queries become a regression test.
We install the widget
A snippet in the theme, with a backup of the theme taken first. From that day on, every search, click and order coming from search is recorded.
When AgentSearch makes sense, and when it doesn't
It's a fit if
- You have thousands or tens of thousands of SKUs and your platform's search returns zero results far too often
- Your customers search by part number, reference or model, and a similar code means a different part
- You sell through several stores from the same catalog, or your reliable stock and product data live in a PIM or an ERP
- You want to know which searches lead to sales, with the data sitting in your own database
You probably don't need it if
- Your catalog is a few hundred well-named products: your platform's search or a plugin such as Doofinder will cost you less
- You need relevance A/B testing or per-customer personalization right now, which AgentSearch doesn't offer yet
Frequently asked questions
Which platforms does AgentSearch work with?
It runs in production on Shopify stores and on top of an ERP, and we have deployed and validated it on PrestaShop 1.7. The widget is dependency-free JavaScript, so it also fits custom builds. We don't have a WooCommerce store in production yet.
Is the AI layer mandatory?
No. It's switched on per store. You can start with BM25, synonyms and facets, and add zero-result suggestions, query interpretation and the chat later. If the AI fails or hits its daily spending cap, search keeps working without it.
Does it work with large catalogs?
Yes. In production it indexes more than 76,000 products and serves several stores from a single index. Keyword search answers in tens of milliseconds on the server, and a full reindex without recomputing vectors takes just over a minute.
Do I have access to my search data?
Yes, and it's stored in your database: queries, zero-result searches, clicks with their position, add-to-carts and attributed orders with their value. The dashboard filters it by store and by country, and personal data is anonymized after 90 days.
Is it an alternative to Algolia or Doofinder?
For large or technical catalogs, yes. Algolia and Doofinder are SaaS products billed by request volume; we deploy and maintain AgentSearch ourselves, with rules specific to your business. For a small, standard store, Doofinder is usually the better choice. The comparison page goes into detail.
Deployment, migration and integration
The work we do around the engine so it performs in your store.
Search for your ecommerce
What it fixes in a store (typos, part codes, facets, per-store stock) and which metrics tell you whether search is selling.
See search for ecommerceMigrating from Algolia or Doofinder
A straight comparison with both, and how to carry over synonyms and rules without leaving your store without search.
See the comparison and migrationWidget and integration
How it installs on Shopify, PrestaShop, WooCommerce or on top of an ERP, and how catalog and stock stay in sync.
See how it integratesWe'll show it to you with your own catalog
Tell us your platform, how many products you have and what search you use today. We'll build the demo with your products.