Learn AI /Search, Embeddings and RAG /Knowledge base
Knowledge base
- Starter
- Search, Embeddings and RAG
- Programme & data teams
Also called: document collection, corpus of record.
Back to the milestone you came from
In five seconds
What it is
An organised set of trusted documents that a search system can look through to find relevant passages for an answer.
In plain words
Think of it as the approved shelf of documents a system looks through before answering a question.
An everyday example
At a food distribution point, the team keeps a shelf of labelled binders: delivery schedules, stock rules, ration tables and site procedures. When a new staff member asks, "How much rice should a household of five receive this month?", the supervisor does not rely on memory alone. They pull the relevant binder, open the right section and read out the current rule. The shelf is not every paper in the building; it is the set of documents the team has chosen to rely on for this job.
A real knowledge base is usually digital, can search thousands of passages in seconds and may use meaning-based matching rather than a person reading tabs by hand. Also, people can judge context more flexibly than a simple search system. The analogy helps with the idea of a trusted reference collection, but it does not show embeddings, ranking or the risk of retrieving the wrong passage.
Remember thisA knowledge base is the curated source collection that retrieval searches. Good indexing helps, but the answer can only be as current, complete and authoritative as the documents included.
Understand it
Understand it
Why it matters
For non-specialists, the knowledge base often determines whether an AI-supported answer is useful, outdated or unsupported. If the collection is incomplete, poorly maintained or includes low-quality documents, the system may retrieve weak evidence and produce misleading output. If the collection is well chosen and current, staff can find relevant passages faster, check sources more easily and make better decisions about whether to trust or reject an answer. In other words, the knowledge base is a governance issue as much as a technical one: what you include, who maintains it and how current it is all shape the quality of the result.
How a knowledge base supports an answer
Show that a knowledge base is the maintained document collection a system searches before returning passages or a grounded answer.
01
1. Collect trusted material
Choose the documents to include, such as policies, manuals, reports or FAQs. The value of the knowledge base depends on what is included and what is left out.
Show a box on the left containing 5 to 7 document icons labelled with examples like policy, manual, report and FAQ. Add a small badge saying selected sources.
02
2. Prepare for search
The documents are often cleaned, split into smaller passages and indexed so the system can search them quickly.
Show the document box feeding into a processing box with three small icons: broom for cleaning, scissors for splitting, and card index for indexing. Use arrows to indicate transformation.
03
3. Match the question to passages
When a question arrives, the system searches the knowledge base to find the passages that seem most relevant by words, meaning or both.
Place a user question bubble above an arrow into a search box. Inside the search box, highlight three passage cards pulled from the indexed collection. Mark them ranked 1, 2 and 3.
04
4. Return evidence or support an answer
The output may be the passages themselves, or those passages may be handed to a language model to help draft an answer with citations.
Show two possible outputs on the right: one panel with highlighted passages, and one panel with a short answer linked by citation markers to those passages.
A left-to-right flow with four boxes. First, selected documents enter the system. Next, the documents are organised for search. Then a user question triggers retrieval of relevant passages. Finally, the system outputs passages, or an answer with citations. Read from left to right to see what goes in, what happens and what comes out.
Hovering or tapping each stage highlights what changes at that step; selecting the final stage toggles between 'passages only' and 'answer with citations'.
Describe this without the diagram
The diagram shows a four-stage process from left to right. Stage one is a selected set of trusted documents, such as policies, manuals, reports and FAQs. Stage two shows those documents being cleaned, split into smaller passages and indexed for search. Stage three adds a user question and shows the system retrieving and ranking several relevant passages from the collection. Stage four shows two possible outputs: either the retrieved passages are shown directly, or they are used to support an answer that includes citations back to the source passages. The point is that the knowledge base is the maintained document collection that search uses before any answer is produced.
With reduced motion: A static four-column diagram replaces animation. Column one shows selected documents. Column two shows preparation steps: cleaning, splitting and indexing. Column three shows a question linked to three ranked passages. Column four shows either passages only or an answer with citations, displayed side by side.
A knowledge base is the collection of material a system searches when it needs source text. It might include policies, manuals, reports, FAQs or approved web pages. The key idea is that the collection is chosen and maintained for a purpose, not gathered at random.
What goes in: documents or passages that an organisation wants to use as reference material. These are often cleaned, split into smaller pieces and indexed. An index is a structure that helps the system find useful parts quickly.
What happens: when someone asks a question, the system searches the knowledge base for passages that seem relevant. In modern systems, this may use semantic search, which looks for similar meaning rather than only matching exact words. The selected passages can then be shown directly or passed to a language model to help draft an answer with sources.
What comes out: ideally, a smaller set of relevant passages, and sometimes an answer grounded in those passages. This is different from training. Putting a document into a knowledge base does not usually change the model's parameters; it gives the system something to look up at answer time.
How “The shelf of official binders at a distribution point” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Shelf of official binders | Knowledge base of selected source documents |
| Labels and sections in binders | Indexing and organisation for fast retrieval |
| Opening the right section for a question | Retrieval of relevant passages |
| Reading the current rule from the binder | Grounded answer based on source text |
A worked example
A reporting officer in a regional bureau is finalising a briefing note for front-office clearance. The approved knowledge base contains current indicator guidance, cleared monitoring reports and notes for the file recording earlier methodological decisions.
Starting from: Question: "Should people who attended two sessions be reported as two participants or one person reached?"
- The source documents have been normalised, divided into passages and indexed by keyword and meaning.
- The retriever searches for passages about attendance, repeat participation and the definition of people reached.
- It returns the relevant indicator definition, a methodological note and a passage from a cleared monitoring report.
- The search tool shows the passages and their sources. The reporting officer checks which guidance is current before revising the briefing note.
Result: The officer uses the current indicator definition, records the methodological choice in the file and cites the source in the briefing note. The knowledge base supports the decision, but does not make it.
In a United Nations workplace
A partnerships officer in a country office is coding open-ended answers from a needs assessment before the coded dataset is uploaded to the portal at close of business. An approved search tool looks through a curated knowledge base containing the coding framework, assessment guidance and cleared notes on earlier coding decisions.
The retriever can find the passages that distinguish similar categories and show the source for each suggestion. The officer still decides how to code the answer, but does not have to reconstruct the methodology from several attachments and the version someone circulated last Tuesday.
Before you try thisThe knowledge base should contain approved guidance, not raw responses by default. Do not upload names, contact details, medical information, protection concerns or other sensitive responses to a tool that has not been authorised for them. Apply the required access controls and retention rules, and have the officer review every coding suggestion.
Where you may hear it
- During a vendor demo for an approved document-search tool
- At an ICT governance board review of sources and access controls
- While clearing terms of reference for a chatbot with citations
- When the front office asks why a search result relied on superseded guidance
Avoid confusion
Avoid confusion
Often confused with
Vector database
A specialised system stores numerical meaning-representations of content and quickly finds the closest matches to a query by semantic similarity.
Training
Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.
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.
Common misconceptions
Common misconception: A knowledge base is the same thing as the AI model's memory.
Correction: Not usually. The knowledge base is an external collection of documents the system can search at answer time. The model is a separate component. Adding a document to the knowledge base does not normally retrain the model or guarantee the model will repeat that document without retrieval.
Common misconception: If a document is in the knowledge base, the answer will automatically be correct.
Correction: Inclusion helps, but it does not guarantee a good result. The system still has to retrieve the right passage, rank it well and use it accurately. Poor chunking, weak indexing, outdated documents or unclear questions can still lead to an unsupported answer.
Common misconception: A vector database and a knowledge base are the same thing.
Correction: They are related but different. A knowledge base is the content collection itself: the documents or passages chosen as sources. A vector database is one possible tool for storing numerical representations of that content to support meaning-based search.
Quick check
Which description best fits a knowledge base in a search or RAG system?
Option 2 states the defining role: a selected source collection searched for relevant passages. Model weights come from training, a bare query history is not necessarily source material, and a rewriting tool changes wording rather than supplying sources.
Go deeper
Go deeper
The simple version
Think of it as the approved shelf of documents a system looks through before answering a question.
The technical version
In retrieval systems, a knowledge base is the curated corpus of source documents used for lookup at inference time. Documents are typically normalised, split into chunks, indexed and sometimes converted into embeddings, which are numerical representations used to compare meaning. A retriever searches this corpus by keywords, semantic similarity or both, and returns candidate passages. Those passages may then be reranked and either shown directly or supplied to a language model in a retrieval-augmented generation pipeline. The knowledge base is therefore the content layer, distinct from the search index, the vector database that may store embeddings, and model training, which changes model parameters rather than providing external source material at query time.
Reviewed external resources
EncyclopaediaKnowledge base (opens in a new tab)Wikipedia
This link gives a broad general overview of the term and suits readers who want a quick starting definition before the AI-specific details.
Intermediatelink checked 2026-07-31
CourseEmbeddings (opens in a new tab)Google
This course resource helps readers understand embeddings, which are often used to search a knowledge base by meaning, and suits beginners who want the technical basics.
Intermediatelink checked 2026-07-31
InteractiveEmbedding Projector (opens in a new tab)TensorFlow
This interactive tool lets readers explore how items can be placed by similarity in vector space, which helps explain semantic search, and suits visual learners.
Starterlink checked 2026-07-31
VideoVideo explanations: Knowledge base (opens in a new tab)YouTube search
This is a YouTube search results page for video explanations of the term, so it suits readers who prefer to compare several teaching styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Knowledge base (opens in a new tab)Reddit search
This is a Reddit search results page for community discussion of the term, so it suits readers who want to see common questions, confusion and practitioner 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
A knowledge base is the curated source collection that retrieval searches. Good indexing helps, but the answer can only be as current, complete and authoritative as the documents included.