Ask your business a question.

What is RAG?

Retrieval-augmented generation, usually shortened to RAG, is a way of answering a question by first retrieving relevant material from sources you control, then generating the answer from that retrieved material. The model still writes the sentence. What changes is that the facts in it came from your documents rather than from the model’s general training.

What is an AI knowledge base?

A set of business sources — policies, handbooks, published pages, help articles, product documentation — that somebody has decided are authoritative, prepared so an AI system can find the specific passage that answers a question. The word "base" does less work than the word "approved": an unreviewed folder is a folder, not a knowledge base.

What is the difference between search and RAG?

Search helps you find documents that mention something. RAG retrieves the relevant passages from those documents and uses them as the context for a written answer, with the passages kept alongside it. Search ends at a list. RAG ends at an answer you can check.

1. 01Source auditWhat exists, who owns it, when it last changed, and whether anyone still stands behind it.
2. 02Authority decidedOne source becomes the one that answers. The decision is yours; we make it explicit.
3. 03The rest excludedOlder and duplicate copies stay in your drive. They stop being answer material.
4. 04Prepared and indexedThe approved source is split into retrievable sections and made available to the system.

1. CustomerPublished material only
2. Support teamPublished, plus internal process
3. Finance teamAll of the above, plus commercial terms

1. Approved FAQsIn scope
2. Warranty PolicyIn scope
3. Priority Support HandbookOut of scope
4. Operations ManualOut of scope
5. Commercial termsOut of scope

The layer inherits the permission model of the system it sits inside rather than inventing a second one, and which reader may reach which source is designed with you. Where a business has no permission model to inherit, defining one is part of the project rather than an assumption inside it.

1. 01The source changesSomeone updates the warranty policy. The old version is still in the drive.
2. 02The change is noticedEither the system watches the source, or a person owns telling it. Which one applies is decided per source, not promised for all of them.
3. 03The old version stops answeringIt is no longer authoritative. It does not need deleting to stop being answer material.
4. 04The new version is preparedSplit, indexed and available to retrieval.
5. 05The same question answers differentlyNobody retrains anything. The evidence changed, so the answer changed.

How is a RAG knowledge base kept up to date?

By changing the source, not by retraining a model. When an approved document or page changes, later answers come from the new version. What has to be designed is how the system finds out — some sources can be watched automatically, others need a person to own the update — and that is decided source by source rather than promised across all of them.

1. An AI chatbotWhat it isThe conversation. How somebody asks, what they see, and how it hands over to a person.Why it is not thisA chatbot needs a knowledge layer to answer from. This is that layer. One is the window, the other is what is behind it.AI Chatbot
2. SearchWhat it isFinding documents that match what you typed.Why it is not thisSearch returns the documents. This returns the answer and keeps the passages attached to it.
3. A vector databaseWhat it isOne possible component for storing and searching passages by meaning rather than by keyword.Why it is not thisA component, not the system. Choosing one answers roughly a tenth of the question. Which sources are authoritative, how they are divided, what the system does when support is weak and who may read what are the other nine tenths. Branditify does not sell a vector database.
4. Fine-tuningWhat it isAdditional training that changes how a model behaves — its style, its format, its handling of a specific task.Why it is not thisIt changes behaviour, not access to current facts. A document that changes next week is a retrieval problem. They are not alternatives and are sometimes used together, but if the goal is "answer from our current documents", retrieval is the direct route.

What is the difference between RAG and a chatbot?

A chatbot is the interface — the conversation a person has. RAG is the knowledge layer that a chatbot can answer from. You can build a chatbot with no approved knowledge behind it, and it will improvise. You can build a knowledge layer with no chatbot at all, and connect it to a phone line or an automation instead.

What is the difference between RAG and a vector database?

A vector database is a piece of infrastructure that can store passages and find them by meaning. RAG is the whole workflow around that: deciding which sources are authoritative, preparing them, retrieving the relevant passages, giving them to the model as context, and controlling what happens when the retrieved material does not support an answer.

Do I need RAG or fine-tuning?

If the problem is that the AI does not know your current business information, retrieval is usually the more direct starting point, because the information keeps changing and retrieval reads the current version. Fine-tuning is normally considered when the problem is how the model behaves rather than what it knows. They are not mutually exclusive, and neither is a substitute for keeping your documents correct.

Answer the same question in three places from three separate copies of your policies and you now have three things to keep correct. That is the failure this arrangement is designed to avoid.

• Custom software When the knowledge layer sits inside a larger system being built.
• AI strategy When the first question is still whether AI is the right answer.

1. Can a priority customer request an expedited replacement?SupportedAnswers, with the three passages attached.
2. And how long does it take?Partly supportedAnswers the supported half. Names the gap.
3. What is the best replacement?Needs clarificationAsks which reading was meant.
4. Can we promise same-day dispatch?Not in approved sourcesSays no approved source carries it.
5. What is the current refund window?Sources disagreeSurfaces the disagreement instead of picking.

The set is written from questions your team is already answering by hand, so it tests the volume you actually have rather than the demo you would like. It is also what makes a change safe later: after a source is added or replaced, the same questions get asked again.

1. 01InventoryWhere knowledge currently lives, who owns each source, and what each system can actually expose.
2. 02Authority mapWhich source answers which kind of question, and which copies are retired from answering.
3. 03PrepareApproved sources cleaned, split into retrievable sections, and checked for the duplicates and contradictions that show up at this stage.
4. 04IndexMade retrievable, with permissions applied.
5. 05TestThe question set is run across every knowledge condition, and retrieval is adjusted against the misses.
6. 06ConnectThe layer is attached to whichever surfaces need it — chat, voice, or an automation.

1. Source conditionWhether one version of each answer already exists, or four do.
2. Number of systemsEvery additional place knowledge lives is another thing to read, and another thing to check permissions against.
3. PermissionsOne audience is a different project from four, and the difference is a business decision before it is a technical one.
4. Update frequencySources that change weekly need a defined path in; sources that never change do not.
5. Surfaces connectedOne channel, or chat and voice and an automation.
6. Question rangeA narrow set of repeated questions is a smaller job than open-ended internal Q&A.
7. LanguagesSource language, question language and answer language are three separate decisions and are not automatically the same.
8. Data sensitivityWhat the sources contain drives where things run and who may reach them.

What determines the scope of a RAG project?

Mostly the state of the knowledge rather than its volume. Ten thousand clean, owned, current documents are a smaller job than four hundred that contradict each other. After that it is the number of systems the knowledge lives in, how many audiences need different views of it, how often it changes, and how many surfaces answer from it.

What we need from you to scope one

The questions your team answers most often, where the answers currently live, and who decides which version is right. That is usually enough for a first shape.

Data and where it runs

What the knowledge layer will hold decides the rest: whether sources contain personal or customer information, which providers may process it, where it may be stored, what is retained, and who may reach it. We confirm those before choosing an architecture rather than fitting them around one. Compliance positions are not claimed on this page — where a certification or a jurisdiction is a requirement, it belongs in scope from the first conversation.

What you own

Your source documents were always yours. So is the configuration — the authority map, the permissions, the question set — along with any code written for you and the accounts the system runs on. There is no Branditify knowledge product to license and no seat price.

1. Detail Driven Car detailing & carewebsite development
2. Lotus Professional Beauty & skincareproduct recommendations · skincare consultation UI
3. Digital Green Agriculture & social impactdata visualisation