Semantic Search
Last updated:
What semantic search is
Traditional keyword search returns documents that contain the words you typed. Semantic search tries to return documents that answer what you meant, even when they use different words. A query for “cheap flights to Rome” can match a page about “low-cost fares to Italy's capital”.
The idea is older than today's language models, but embeddings made it practical for everyday applications such as site search, recommendations and AI chatbots.
How it works
Most semantic search systems follow the same steps:
- Content is split into passages, and each passage is converted into an embedding by a model.
- The query is converted into an embedding with the same model.
- The system measures the distance between the query vector and the stored vectors, often with cosine similarity.
- The closest passages are returned, ranked by similarity.
An example
A software company's help center has an article titled “Resetting your login credentials”. A customer types “I forgot my password and can't get in”. A keyword search might rank the article low, because only one word overlaps. Semantic search recognizes that both texts describe the same problem and returns the article first.
The reverse also happens. A query for the error code “E-4012” is about exact characters, not meaning, and semantic search may return articles about other error codes that look similar to the model. Because similarity is a matter of degree, semantic search also always returns its best candidates, even when nothing truly relevant exists, so good systems add further checks before using the results.
Why it matters for business chatbots
Visitors rarely use the same vocabulary as your documentation. They write short, informal questions, make typos or describe a symptom instead of naming a feature. Semantic search lets a chatbot find the right passage anyway, which is the basis for answers grounded in your own content rather than in the model's general knowledge.
Its blind spots are exact terms: SKUs, part numbers, names and acronyms. For that reason, production chatbots usually pair semantic search with keyword search, an approach known as hybrid search.
How intoCHAT uses it
intoCHAT creates an embedding for every chunk of your training content and runs a semantic search for each visitor question. It combines those results with a keyword search, so the agent finds passages both by meaning and by exact terms. Q&A pairs you write take priority, and sources you star rank higher, without retraining.