Semantic search
Semantic search: by meaning,not by keywords.
Semantic search understands the intent behind the words and returns the right results even if the term does not appear exactly in the text. It works in your online store, catalogue or document archive, with no conversation and no generated answers: just a better list of results.
“No results found”
A customer searches for “trainers for running on tarmac” and gets nothing, because the product is listed as “road running footwear”. An employee searches for “late payment penalties” and misses the contract that talks about “liquidated damages”. The information exists, but the search does not make the connection.
- Searches with no results for products or documents that do exist
- Synonyms, missing diacritics and typos treated as different terms
- Results ranked by how often a word appears, not by relevance
- Users who give up or call support instead of finding things themselves
Who it is for
- Online stores with large catalogues, where customers do not know a product's technical name
- Companies with document archives where employees search by idea, not by title
- Websites with a lot of content: articles, guides, technical data sheets, FAQs
- Teams that want better results without the risk of AI-written answers
What semantic search does
- 01
Search by intent
Interprets what the person is looking for and finds products or documents close in meaning, even with no words in common.
- 02
Hybrid search
Combines exact matching — codes, SKUs, proper names — with matching by meaning, so neither is lost.
- 03
Tolerance of phrasing
Recognises synonyms, missing diacritics, typos and colloquial terms.
- 04
Filters still work
The usual filters — price, category, date, department — work on top of the semantic results.
- 05
Similar results
Suggests products or documents similar to the one currently open.
- 06
Failed search report
Shows what people search for and cannot find, so the team can fill gaps in the catalogue or documentation.
How it works
How semantic search works
The steps marked in yellow are approved by a person.
- Input
- AI
- Human in the loop
- Action
- Result
- 01
Input
Catalogue or documents
Products or documents are imported with titles, descriptions and attributes.
- 02
AI
Indexing by meaning
Each item gets a representation of its meaning, alongside the classic index.
- 03
Human in the loop
Validated rules
The team validates synonyms, promoted products and result ranking on real searches.
- 04
Action
Hybrid search
For each search, the index is queried by meaning and by exact words, with filters and permissions applied.
- 05
Result
Ranked results
The user gets the list of results, with no AI-generated text.
What data semantic search uses
- The product catalogue or document collection, with titles, descriptions and attributes
- The search history from the website or application, if available
- Lists of synonyms, specialist terms and product codes
- Business rules: stock, promoted products, restricted documents
- A set of real searches with expected results, for testing
Possible integrations
- WooCommerce
- Shopify
- Magento
- PrestaShop
- Elasticsearch
- OpenSearch
- Algolia
- PostgreSQL with pgvector
- Google Drive
- Microsoft SharePoint
- WooCommerce
- Shopify
- Magento
- PrestaShop
- Elasticsearch
- OpenSearch
- Algolia
- PostgreSQL with pgvector
- Google Drive
- Microsoft SharePoint
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What people validate
- Synonyms and ranking rules, before publication
- Products promoted in or excluded from the results
- Regular review of searches with no results or weak results
- Restricted documents in internal search and who has access to them
Limitations and how we handle them
- Search by meaning returns close results, not necessarily exact ones; for codes and SKUs exact matching matters, which is why we use hybrid search
- Results are better the more complete the catalogue's descriptions and attributes are
- Highly specialised terms or internal jargon may require synonym lists maintained by the team
- It does not provide answers or explanations; if you need answers written from documents, you need a RAG system
How we measure search quality
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Zero-result searches
- The share of searches that return nothing, compared before and after launch, from the search logs.
- Relevance of top results
- On a set of real searches, the team checks whether the expected results appear near the top.
- Interaction with results
- How often users open a result and what position it holds in the list.
- Response time
- The duration of each search, measured continuously, so the experience stays fast.
How we implement semantic search
- 01
Analysing current searches
We study what users search for and where the current search fails.
- 02
Preparing the data
We clean up descriptions and attributes and build the hybrid index.
- 03
Comparative testing
We compare the new results with the old ones on real searches, together with your team.
- 04
Launch and tuning
We launch gradually and adjust synonyms and rules based on the failed search report.
Use cases
These scenarios are conceptual examples showing what can be built. They are not delivered implementations or results achieved for clients.
- Online car parts store
Search by symptom or everyday name
The customer searches for “front brake pads Logan 2015” or “squeaks when braking”, and the search returns the right products, with the model filters kept.
Conceptual example - Building materials distributor
B2B technical catalogue
Trade customers find products by application, for example “adhesive for outdoor tiles”, not only by code.
Conceptual example - Law firm
Search in the internal archive
Lawyers find documents by concept, and the results respect each person's access to case files.
Conceptual example
Frequently asked questions
What is the difference from a RAG system?
Semantic search returns a list of results, which the user opens and reads. A RAG system goes further and writes an answer from them. Search without chat removes the risk of a wrongly generated answer and is often enough for stores and archives.
Does searching by product code still work?
Yes. Hybrid search keeps exact matching for codes, SKUs and proper names and combines it with matching by meaning.
What happens when the search finds nothing suitable?
It shows the closest results and suggestions for rephrasing, and the search goes into the failed search report. It does not invent results or generate text.
What data is used and who sees the results?
Only the catalogue or documents you choose are indexed. Search history can be kept without identifying data, and in internal search the results respect each user's access rights. The hosting location is decided according to your requirements.
Does it work well in Romanian?
The models that represent the meaning of text support Romanian, but quality varies between them; we test them on your searches and terms before choosing.
What is hybrid search?
It is the combination of searching by exact words and searching by meaning. Product codes, SKUs and proper names are found exactly, while free-form phrasing is understood by intent, in a single list of results.
Can’t find the answer? Write to us and we’ll reply directly.
Want search that understands what your customers are looking for?
Send us a few searches that return no results today. We will show you how semantic search would handle them and how we measure its effect.
Related solutions
- RAG systemsAssistants that search the configured sources before answering and show where each piece of information comes from.
- AI assistant for e-commerceGuides shoppers to the right product using your catalogue, answers questions about availability and orders, and prepares cart recovery messages only where consent exists.
- Enterprise knowledge basesThe infrastructure that keeps company knowledge in order: inventoried sources, permissions, updates and an owner for each domain.