RAG systems
RAG systems: first they search,then they answer.
A RAG (retrieval-augmented generation) system is an AI assistant that, before answering, searches the configured data sources and builds its answer on the passages it finds, with references to them. It improves relevance and makes answers verifiable, although it does not eliminate every error.
A language model does not know your documents
An LLM answers from what it learned during training: it does not know your procedures, prices or contracts, and it may fill the gaps with plausible but false information. Loading every document into a conversation is impractical and does not respect access rights.
- Confident answers with no source that can be checked
- Outdated information, because the model does not know which document has been replaced
- The risk that a user receives excerpts from documents they have no access to
- No rule for questions the documents do not answer
Who it is for
- Companies that want an AI assistant that answers from their own documents, not from general knowledge
- Teams that need to be able to check where every answer comes from: quality, compliance, legal
- Companies that have tried a chatbot on their documents and got invented or unsourced answers
- Technical teams that want to understand and control how the system chooses its information
What a RAG system does, step by step
- 01
Preparing the sources
Documents are split into chunks and indexed by meaning, keeping the date, version and access level.
- 02
Finding relevant passages
For each question, the system looks for the chunks closest in meaning, only from the sources the user is allowed to see.
- 03
Answer from the chunks
The model receives the question and the passages found, and is instructed to answer only on that basis.
- 04
Verifiable citations
The answer points to the document and passage used, so the user can open the source.
- 05
Reasoned refusal
If the passages found do not contain the answer, the system says so instead of making something up.
- 06
Reindexing on change
Modified or withdrawn documents are updated in the index on an agreed schedule.
How it works
A question's journey through a RAG system
The steps marked in yellow are approved by a person.
- Input
- Action
- AI
- Human in the loop
- Result
- 01
Input
The user's question
Your colleague asks in their own words, in a chat or in an application.
- 02
Action
Access filtering
The system identifies the user and limits the search to the sources they are entitled to.
- 03
Action
Index search
The chunks most relevant in meaning are found, from current documents.
- 04
AI
Answer with citations
The model writes the answer from the chunks or says the information is missing.
- 05
Human in the loop
Checking the source
The user opens the citation and confirms the information before relying on it.
- 06
Result
Answer and feedback
The answer is used, and the user's rating feeds into the quality analysis.
What components a RAG system has
- The configured sources: documents, internal pages, resolved tickets, tables
- Metadata for each document: date, version, department, access level
- A semantic search index (embeddings), updated regularly
- Instructions for the model: answer only from the sources, cite them, acknowledge missing information
- A set of test questions with known correct answers
Possible integrations
- Google Drive
- Microsoft SharePoint
- Confluence
- Notion
- PostgreSQL with pgvector
- Qdrant
- Elasticsearch
- PDF and Word files
- OpenAI API
- Anthropic API
- Google Drive
- Microsoft SharePoint
- Confluence
- Notion
- PostgreSQL with pgvector
- Qdrant
- Elasticsearch
- PDF and Word files
- OpenAI API
- Anthropic API
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What people validate
- Choosing which sources go into the index and withdrawing outdated ones
- Access rights, aligned with the permissions in the original systems
- Answers used for important decisions, checked directly against the source
- Regular review of answers marked as wrong and of unanswered questions
Limitations and how we handle them
- RAG reduces the risk of invented answers but does not eliminate it: the model can misinterpret a correct passage
- If the search returns the wrong chunk, the answer will be wrong, even if it has a citation
- Scanned documents, complex tables or images may be read incompletely
- Contradictory information from different sources is not resolved automatically; the system can only flag it
- Answers are only as current as the last reindexing of the sources
How we evaluate a RAG system's quality
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Relevance of retrieved chunks
- Whether the chunks retrieved for each test question contain the answer — checked by the team.
- Faithfulness to sources
- Whether the answer says only what is in the chunks, assessed on a sample of real answers.
- Citation accuracy
- Whether the cited document actually supports the statement, checked manually by spot checks.
- Correct refusals
- Behaviour on questions the sources do not answer, deliberately included in the test set.
- Index freshness
- The time between a document being changed and the new version being used in answers.
How we implement a RAG system
- 01
Sources and access
We agree which sources are included, who is entitled to what and how often the documents change.
- 02
The test set
We gather real questions with known answers, including questions the documents do not answer.
- 03
Indexing and tuning
We split and index the documents, then tune the search until the chunks retrieved are relevant.
- 04
Pilot and reindexing
We launch to a small group, analyse the feedback and set the schedule for updating the sources.
Use cases
These scenarios are conceptual examples showing what can be built. They are not delivered implementations or results achieved for clients.
- Industrial equipment manufacturer
Answers from technical data sheets
Sales engineers ask about specifications and get the answer with the relevant data sheet page cited, which they check before sending it to the customer.
Conceptual example - Insurance broker
Policy terms
Advisers ask what a product covers, and the system answers from the terms in force, citing the clause; explaining it to the customer remains the adviser's job.
Conceptual example - Private clinic network
Internal administrative procedures
Reception staff find the steps for reimbursements or rescheduling in the approved procedures, without the system having access to patient data.
Conceptual example
Frequently asked questions
What does RAG mean?
RAG stands for retrieval-augmented generation, meaning generation supported by search. The system first searches your documents for relevant passages, then the language model writes the answer based on them and shows the source.
How is RAG different from an ordinary chatbot?
An ordinary chatbot answers from the model's general knowledge. A RAG system searches your sources first and answers on that basis, with a citation. The practical difference: the answer can be checked.
Does RAG eliminate mistakes?
No. It reduces them and makes them easier to spot. The system may pick the wrong passage or misinterpret a text, which is why we show the sources, test on real questions and recommend human review for important decisions.
What happens when the documents do not contain the answer?
The system is instructed to say clearly that it did not find the information and, where appropriate, to point the user to someone who can help. These questions are logged, because they show what is missing from the documentation.
Who sees what in the answers?
The search is filtered by the user's permissions before the model receives any passage. Only the relevant passages reach the model provider, not the whole archive; for sensitive data we also consider privately hosted models, and the processing of personal data is assessed in line with the GDPR.
How is it different from an enterprise knowledge base?
This page describes the technique: how the system searches and answers. An enterprise knowledge base deals with the surrounding infrastructure: which sources exist, who owns them, how they are updated and how they are governed across the whole company.
How often are the sources updated?
It depends on the pace of change: on every change for critical documents, daily or weekly for the rest. Withdrawn documents are removed from the index so they are no longer cited.
Can’t find the answer? Write to us and we’ll reply directly.
Want AI answers you can check against the source?
Send us a few types of documents and the questions that come up often. We will explain what the RAG system would look like and how we test its quality.
Related solutions
- Enterprise knowledge basesThe infrastructure that keeps company knowledge in order: inventoried sources, permissions, updates and an owner for each domain.
- Semantic searchSearch by meaning across your website, catalogue or documents: relevant results even when users do not type the exact words.
- AI evaluation and monitoringTest sets, checks at every change, human spot-check reviews and cost tracking, so you know how your AI performs over time.