Vector databases and embeddings help e-commerce brands match products to what shoppers mean, not just the words they type. An embedding turns a product title, description or image into a list of numbers that captures meaning; a vector database stores those lists and finds the closest matches quickly. The practical results are better on-site search, related-product suggestions and faster catalogue clean-up.
Keyword search fails in familiar ways. A shopper types “summer dress for a beach wedding” and gets nothing, because no product contains that exact phrase. Embeddings let the store recognise that a light linen midi dress is a good answer. That is the core idea, and most use cases grow from it.
How do embeddings and vector databases work together?
An embedding model, available from providers such as OpenAI, Google, Cohere or as open-source models, converts text or images into vectors with hundreds or thousands of dimensions. Items with similar meaning end up close together in that space. A query is embedded the same way, and the system looks for the nearest product vectors.
The vector database handles storage and fast nearest-neighbour search at scale. Options include managed services such as Pinecone, open-source engines such as Weaviate, Qdrant and Milvus, and extensions like pgvector for PostgreSQL. Elasticsearch and OpenSearch also support vector fields, which helps if you already run them for search.
Which e-commerce use cases pay off first?
Search is usually the first win, because failed searches are visible in analytics and directly cost sales. Recommendations and catalogue work follow.
- Semantic search: understand descriptive queries, synonyms and misspellings beyond what a synonym list covers.
- Similar products: show visually or functionally similar items when a product is out of stock.
- Image search: let shoppers upload a photo and find matching styles, using image embeddings.
- Duplicate detection: find near-identical listings created by different suppliers or imports.
- Auto-categorisation: suggest a category for new products based on similar existing ones.
- Support answers: retrieve the right return policy or size guide paragraph for a chatbot.
Should you replace keyword search with vectors?
Usually not. Hybrid search, which combines keyword relevance (such as BM25) with vector similarity, tends to beat either alone. Keywords stay strong for exact matches like model numbers, sizes and brand names; vectors help with descriptive and fuzzy queries.
Many commerce search platforms, including Algolia, Elastic and Shopify’s own search features, now blend both approaches. Check what your current platform offers before building a custom stack.
Ranking still needs business rules
Similarity is not the only goal. Stock levels, margin, delivery time and seasonality still matter. Apply those as filters or re-ranking factors after retrieval, so a perfect but out-of-stock match does not top the results.
What does it cost to run?
Costs come from creating embeddings, storing vectors and serving queries. Embedding a catalogue of tens of thousands of products with a hosted model is typically modest, since each product only needs embedding once and again when it changes. Query costs scale with traffic, and storage scales with catalogue size and vector dimensions.
Engineering time is the larger cost. Someone has to build the pipeline that keeps vectors in sync with the catalogue, tune hybrid ranking and monitor quality. A small store on a hosted platform may get most of the benefit from built-in features with no engineering at all.
How do you measure search quality?
Start from data you already have. Analytics tools record site search terms, and many platforms report zero-result searches. Those are your baseline problems.
- Zero-result rate: share of searches returning nothing.
- Search exit rate: share of searchers who leave right after searching.
- Search conversion: orders from sessions that used search.
- Click position: how far down shoppers scroll before clicking.
- Manual relevance checks: a panel of 50 real queries rated before and after changes.
What are the common pitfalls?
Poor product data undermines everything. If descriptions are thin or copied from suppliers, embeddings have little to work with. Enrich titles and attributes first. Another pitfall is ignoring language: multilingual stores need an embedding model that handles all their languages well, or separate indexes per language.
Freshness is the third. When prices, stock or descriptions change, vectors and metadata must update too. Build sync into the product update process instead of running occasional bulk jobs.
How long does a first project take?
For a store on a platform with built-in semantic search, configuration and testing can fit into a couple of weeks. Most of that time goes into reviewing query logs, adjusting synonyms and checking results by hand. A custom stack takes longer, because you need an embedding pipeline, a sync process for catalogue updates, hybrid ranking and an evaluation set.
Plan the custom route in phases. Phase one handles search only, for one language and one main category. Phase two adds similar-product suggestions and the remaining categories. Phase three covers image search or support retrieval if the data supports it. Each phase should end with a relevance check against the same query panel, so you know the next phase is building on something that works.
A useful rule is to ship phase one to a share of traffic, for example through an A/B test on the search page, rather than to everyone at once. If search conversion drops, you can roll back without losing a season of sales.
How Alien Road applies vector databases for e-commerce clients
In a typical engagement, discovery starts with the client’s search logs: top queries, zero-result queries and search conversion. We also review product data quality, because enrichment often comes before any vector work.
Setup depends on the platform. For stores on hosted platforms, we start by configuring built-in semantic or hybrid search and synonym rules. For custom stacks, we build an embedding pipeline, choose a vector store that fits existing infrastructure and add business-rule re-ranking. Measurement compares zero-result rate, search conversion and a fixed relevance panel before and after. Reporting is monthly, with a list of failing queries to fix next.
Where to start this week
Export last month’s site search terms and sort by frequency. Mark which of the top 100 return no results or obviously wrong ones. If many are descriptive phrases rather than typos, semantic or hybrid search is likely worth testing; if most are misspellings, better spelling tolerance may be enough.
Keep reading: AI Optimization · AI Optimization