Learn AI /Search, Embeddings and RAG /Semantic search
Semantic search
- Starter
- Search, Embeddings and RAG
- Programme & data teams
- Policy officers
Also called: vector search, meaning-based search.
Back to the milestone you came from
In five seconds
What it is
A search method that finds material by similar meaning, not just shared words, and returns the most relevant passages for review or further AI use.
In plain words
It is a way to search by meaning. Instead of looking only for the same words, it tries to find text that talks about the same idea.
An everyday example
You ask for something that will stop a cupboard door from slamming. The packages are labelled “soft-close damper”, “cabinet buffer” and “hinge cushion”, none of which repeats your wording. A helpful assistant considers what you are trying to achieve and brings back the most relevant options. Semantic search works similarly: it matches the intended meaning, then presents likely results for you to inspect.
A librarian uses broad world knowledge, judgement and context in ways a search system does not fully replicate. Semantic search relies on numerical representations learned from data, so it can miss rare terms, legal wording, names or codes that a human librarian might notice immediately. It also does not by itself verify whether a retrieved passage is correct or appropriate for the task.
Remember thisSemantic search retrieves passages with similar meaning, even when they use different words. The results still require review in their original context.
Understand it
Understand it
Why it matters
People rarely search using the exact words found in a document. That matters in policy, programme and public information work, where the right source may use different terminology, abbreviations or phrasing. Semantic search can help users find the relevant passage faster, which supports better decisions and more transparent answers. It is also useful before an AI system drafts a response, because retrieving relevant source text can reduce unsupported output. But it is not magic: if the index is poor, the passages are badly split, or the topic needs exact legal wording, results can still be wrong or incomplete. Non-specialists should care because search quality shapes what evidence is seen first.
How semantic search finds passages by meaning
After viewing this, a reader should be able to describe how a query and stored passages are turned into vectors, compared for similarity, and returned as ranked results.
01
1. User asks a question
The search starts with a short query, such as “guidance on drought-related crop loss”.
Draw a query box on the left with sample text inside and an arrow pointing right.
02
2. Documents are prepared
Documents are split into smaller passages so the system can return the specific part that matches.
Above or below the query path, show a stack of documents breaking into several smaller passage cards.
03
3. Text becomes vectors
The query and each passage are converted into embeddings, which are number patterns intended to place similar meanings near each other.
Show both query and passage cards passing into matching encoder boxes, then into small dot clusters or compact number bars labelled vectors.
04
4. Similarity is measured
The system compares the query vector with stored passage vectors and looks for the nearest matches in meaning.
Show the query vector entering a comparison area with several passage vectors; highlight three nearest ones with stronger colour or outline.
05
5. Best passages are returned
The search outputs a ranked list of the most relevant passages, which a person or another system can inspect.
On the right, show a ranked list labelled 1, 2, 3 with short passage snippets and source document labels.
A left-to-right flow with two inputs: a user query and a document collection split into passages. Both are converted into vectors, compared in a similarity step, then the closest passages are returned as ranked results. Read from left to right to follow the process from input to output.
Hovering or tapping a step highlights the matching objects and arrows for that step; a toggle switches the example query text to another wording to show that similar-meaning results can still appear.
Describe this without the diagram
The diagram explains semantic search as a five-step flow. First, a user enters a question. Second, documents are split into smaller passages. Third, both the question and the passages are converted into embeddings, which are numerical vectors intended to place similar meanings close together. Fourth, the system compares the question vector with the stored passage vectors and identifies the nearest ones by similarity. Fifth, it returns a ranked list of passages. The main lesson is that matching is based on similar meaning, not only shared words.
With reduced motion: A static left-to-right diagram shows the same five stages with numbered boxes and arrows. The three best-matching passages are highlighted without animation, and the alternate-query example appears as a small side note instead of a transition.
Ordinary search often looks for exact words or close matches. Semantic search tries to find text with a similar meaning even when the wording is different. For example, a search for “flood damage to roads” may also find a passage about “washed-out transport routes after heavy rain”.
What goes in? Usually a user question and a collection of documents. The system turns the question and document passages into numerical representations called embeddings. An embedding is a list of numbers designed so that items with similar meaning end up closer together.
What happens next? The system compares the question’s embedding with the embeddings of many stored passages and retrieves the nearest ones. “Nearest” here does not mean physically nearby. It means most similar in that numerical space. The results may then be reordered by another ranking step.
What comes out? A list of passages, documents or records that are likely to match the user’s intent. This is different from retrieval-augmented generation: semantic search finds the source material, while retrieval-augmented generation uses retrieved material to help a model produce an answer. Semantic search can help across different phrasing or terminology, and across languages when its embedding model supports them, but it still needs testing because it can miss important exact terms or return items that only seem related.
How “Finding the right fitting at a hardware counter” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The student’s question in everyday language | The user query |
| The librarian focusing on meaning rather than exact wording | Semantic matching using embeddings |
| Nearby shelves on related topics | Items close together in embedding space |
| The books brought back first | Top retrieved search results |
A worked example
A programme officer in a regional bureau reviews approved monitoring-report passages while preparing a progress report for clearance. The officer searches for “delivery slowed because the partner changed staff”. One passage mentions personnel turnover, another timely invoices, and a third training for newly appointed personnel.
Starting from: Query: “delivery slowed because the partner changed staff”
- The system divides the approved reports into passages and stores an embedding for each passage.
- It converts the officer’s query into an embedding using the same model.
- It compares the query embedding with the stored passage embeddings and identifies the nearest passages by semantic similarity.
- The passage stating that “delivery slowed following turnover among implementing partner personnel” ranks first, despite sharing few exact words with the query.
- The passage about training newly appointed personnel may rank next because it is related, while the passage about timely invoices ranks lower. Filters or reranking can refine the order.
Result: The officer receives the passage about personnel turnover first and reviews it in the original monitoring report before using it. The system has retrieved a meaning match, not confirmed a cause or supplied cleared wording automatically.
In a United Nations workplace
A national officer in a Headquarters department is finalising a background note at close of business, with several documents already waiting for translation. To find earlier wording for “activities paused because contributions arrived late”, the officer searches an approved repository of cleared passages. Semantic search surfaces “implementation was delayed pending receipt of funds”, although the wording differs substantially.
The search compares the meaning represented by the query and passage embeddings, rather than requiring exact terms. It can quickly surface precedent wording for the officer and translators to review, but it does not determine which formulation is authoritative or suitable in the new context.
Before you try thisDraft text containing personal, protection-sensitive, security-sensitive or other confidential information must not be entered into a search tool unless that tool is explicitly approved for the information and purpose. Retrieved passages must also be checked in their original context and against the latest cleared text.
Where you may hear it
- A vendor demo during an ICT governance board meeting
- A first reading of requirements for an internal document search tool
- A knowledge management discussion in a regional bureau
- A briefing on retrieval-augmented generation before an approved collection is indexed
Avoid confusion
Avoid confusion
Often confused with
Keyword search
A search method that finds documents by matching query words to indexed text and ranking the matches, often with scoring methods such as BM25.
Retrieval-augmented generation
A method that first finds relevant source passages and then asks a language model to answer using that material.
Indexing
Indexing organises documents or document chunks into a searchable structure so a system can quickly find the most relevant passages for a query.
Side-by-side
Common misconceptions
Common misconception: Semantic search replaces keyword search completely.
Correction: Not always. Meaning-based matching is helpful when wording varies, but exact keywords still matter for names, article numbers, codes, legal clauses and rare technical terms. Many practical systems combine both approaches.
Common misconception: If two texts are close in semantic search, they must say the same thing.
Correction: Close vectors usually suggest related meaning, not identical meaning. Two passages can be topically similar but differ in details, time period, scope or factual accuracy, so results still need checking.
Common misconception: Semantic search produces final answers.
Correction: By itself, it usually retrieves relevant passages or documents. Another step is needed if a system will summarise or generate an answer, and that later step can still make mistakes.
Quick check
Which statement best describes semantic search?
Semantic search retrieves text close in meaning, often using embeddings. Word matching is keyword search. Organising searchable passages is indexing. Retrieving passages before generating an answer is retrieval-augmented generation.
Go deeper
Go deeper
The simple version
It is a way to search by meaning. Instead of looking only for the same words, it tries to find text that talks about the same idea.
The technical version
In a typical setup, the query and document passages are encoded into embeddings: dense numerical vectors trained so semantically similar text tends to lie closer together in vector space. At search time, the system computes the query embedding, compares it against stored passage embeddings using a similarity measure such as cosine similarity, and retrieves the nearest neighbours. Results may then be filtered, reranked or combined with keyword methods. This is different from indexing in general, from the vector database that may store the vectors, and from retrieval-augmented generation, which uses retrieved passages to condition a model’s output.
Reviewed external resources
EncyclopaediaSemantic search (opens in a new tab)Wikipedia
This link gives a broad overview of the term, its history and related ideas, and suits readers who want a general reference starting point.
Intermediatelink checked 2026-07-31
CourseEmbeddings (opens in a new tab)Google
This link explains embeddings, which are a core building block of semantic search, and suits beginners who want a guided course-style introduction.
Intermediatelink checked 2026-07-31
InteractiveEmbedding Projector (opens in a new tab)TensorFlow
This interactive tool lets readers explore how embeddings can be positioned and compared, and suits visual learners who want hands-on intuition.
Starterlink checked 2026-07-31
VideoVideo explanations: Semantic search (opens in a new tab)YouTube search
This is a YouTube search results page for video explanations of semantic search, and it suits readers who prefer to compare several video styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Semantic search (opens in a new tab)Reddit search
This is a Reddit search results page for community discussion about semantic search, and it suits readers who want to see common questions, practical experiences and differing viewpoints.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Semantic search retrieves passages with similar meaning, even when they use different words. The results still require review in their original context.