Embeddings and semantic search explained for PMs
An embedding is a list of numbers, a vector, that represents the meaning of a piece of text or an image, so that items with similar meanings get vectors close to each other. Semantic search converts the query into an embedding and retrieves the stored passages whose vectors are nearest, often measured by cosine similarity. It finds a page about how to “end your plan” when the user typed “cancel my subscription”.
Definition written by Alex Dionisio, Senior Product Manager. Updated November 13, 2026.
Why it matters for a PM
Embeddings sit behind RAG assistants, similar-item recommendations, duplicate ticket detection and search bars that understand intent. A PM can skip the math, yet should understand what drives relevance: how documents are split, which embedding model is used, how results are filtered and ranked, and how retrieval quality is measured on real queries.
Example
In an internal wiki, an employee searches for “parental leave”. Keyword search finds nothing because the page is titled “Family policy”. Semantic search returns it first, since the two phrases have close embeddings, and a filter limits results to pages from the employee's own country.
How does an embedding work?
An embedding model takes a piece of text, an image or other content and returns a list of numbers, often a few hundred or a few thousand of them. Those numbers place the content in a space where distance means something: two texts about the same topic end up close together, even when they use different words. Closeness is most often computed with cosine similarity.
The model learned those positions from huge numbers of related text pairs, such as a question and its answer. Its sense of meaning therefore mirrors the material it was trained on: niche industry jargon or an underrepresented language may be placed less accurately, which is why testing on your own content matters.
Embeddings are not tied to a single language either: multilingual models place a French question close to the English passage that answers it, which can let one index serve several markets, provided you test that pairing on your own content.
Keyword, semantic or hybrid search: which should you choose?
Keyword search excels at exact terms: an order number, an error code, a product name. Semantic search understands rephrasing and synonyms, but can blur two neighboring concepts. In most products, combining both gives the best results.
The right mix depends on real queries. Before deciding, export a few hundred searches from your users and look at their shape: precise references argue for keywords, everyday questions argue for semantic search. Watch the zero-result and low-click queries too: they reveal the gaps between how your users talk and how your content is written.
How do you evaluate embedding-based search?
Collect real queries and mark, for each one, the documents that should come up. Then measure the share of queries where a good document appears in the top results, and the position of the first good document. Compare several embedding models and several chunk sizes on that same set: the gaps are often larger than expected. Keep the set small enough to review by hand but broad enough to cover your main query types.
Think about the lifecycle of the index too: every newly published document has to be embedded quickly, and every deleted document, or one whose permissions change, has to drop out of the results. An index that lags behind its source serves outdated answers even after the page has been fixed.
What are embeddings used for beyond search?
The same representation can group customer feedback by theme automatically, flag duplicates, recommend content similar to what a user has viewed, or detect an unusual message. For a PM it is an analysis tool in its own right: clustering thousands of verbatims by similarity gives a first map of the topics, which a human then reviews and names. The same vectors also help catch duplicate content before it goes live, which is useful in a help center where several authors cover overlapping topics.
Key points
- Vectors live in a vector database or index, usually alongside metadata that allows filtering.
- Semantic search can miss exact terms such as product codes or names; hybrid search adds keyword matching such as BM25.
- A reranking step can reorder the top results with a more precise model.
- Changing the embedding model means computing the vectors of the whole corpus again.
Common mistakes
- Returning the nearest results even when none is relevant, instead of applying a threshold.
- Splitting documents into chunks too small to carry meaning, or too large to be specific.
- Assuming access rights are respected because the search is fast and accurate.
Go further with Module
The courses and lessons that cover this concept:
Guide on this topic
- RAG vs fine-tuning: how to choose
What each approach solves, the decision criteria and how to evaluate, with a decision grid.