Enterprise search beyond keywords
Why pure vector search is not enough for enterprise commerce, and how keyword precision, semantic understanding, business rules and product knowledge combine into hybrid discovery.
- Type of work
- Perspective
- Context
- Enterprise commerce · Perspective
- Disciplines
- Search, Semantic retrieval, Evaluation, Merchandising
Problem
Language models made semantic search easy to build, and it is tempting to replace keyword search with vectors. In commerce that is rarely a good trade. Semantic retrieval understands “a warm jacket for cycling to work”, then misses the exact size, article number or technical standard the next customer types.
Context
Enterprise catalogues are not just text. Products carry structured attributes such as sizes, materials, standards and compatibility. Search results carry business intent too: availability, assortment and merchandising priorities.
Customers search in two modes. Sometimes they are exact: an article number, a size, a specification. Sometimes they describe a need in their own words. Good search handles both, often within the same query.
Why it mattered
Search is where customers state their intent most directly. Precision failures show the wrong product. Recall failures hide the right one. Ignoring business rules breaks merchandising, or promises products that cannot be delivered.
Pure vector search improves understanding but gives up precision, explainability and control. Pure keyword search keeps control but fails on anything phrased in the customer’s own words.
Approach
Treat discovery as a layered system in which each layer answers a different question:
- Keyword retrieval for precision on exact terms, codes and attributes.
- Semantic retrieval for meaning, synonyms and descriptive queries.
- Business rules for merchandising, availability and assortment.
- Product knowledge for structured attributes and the relationships between products.
Ranking combines the layers. Evaluation keeps the combination honest: relevance judgements, precision and recall, and the commercial outcomes search is meant to move.
System model
-
Keyword search
- Answers
- Exact terms, codes and attributes
- Example
- “W32 L34”
- On its own
- Misses intent and synonyms
-
Semantic retrieval
- Answers
- Meaning and descriptive language
- Example
- “warm jacket for cycling to work”
- On its own
- Blurs exact attributes
-
Business rules
- Answers
- Availability, assortment and merchandising
- Example
- Only what can be delivered
- On its own
- Invisible if buried in a model
-
Product knowledge
- Answers
- Structured attributes and relationships
- Example
- What else the job needs
- On its own
- Only as good as the product data
Hybrid product discovery
Ranking combines the layers; evaluation ties relevance to commercial outcomes.
Decisions
-
Hybrid by default
Combine lexical and semantic retrieval and tune how they are weighted, instead of replacing one with the other.
-
Keep structured attributes first-class
Filters, facets and attribute matching stay precise and explainable. An embedding should never have to approximate a size.
-
Make business rules explicit
Merchandising, availability and assortment rules live in a layer the business can see and change, not inside an embedding.
-
Evaluate relevance and commercial impact
Judge search on precision, recall and the commercial outcomes it drives, not on how impressive a demo query looks.
What it changes
The question moves from “keyword or vector?” to “which layer should answer this query, and how do we know it worked?”
- Precision where customers are exact, understanding where they are not
- Business rules that stay visible and controllable
- Evaluation tied to commercial outcomes rather than demo quality
What I learned
Relevance is a business decision as much as a technical one.
A perspective drawn from commerce search work. No client data or performance figures are shown.