Cited
Every answer
The exact passage and source document appear with the answer. Verifying takes five seconds.
AI MODULE 03
AI knowledge base (RAG)
An AI that answers from your documents, not from what it read online.
RAG — retrieval-augmented generation — forces the model to search your documents before answering, then cite what it used. That is the difference between an AI that guesses and an AI that retrieves. We build that base, connect it to your sources and keep it current as your documents change.
Who it is for
- Organisations sitting on thousands of pages: quotes, contracts, standards, spec sheets
- Consultancies and expert practices whose value rests on precision
- Manufacturers with catalogues and product sheets that change often
- Any team that tried a generic chatbot and dropped it for lack of reliability
What you live with today
- Consumer AI invents product references that do not exist in your catalogue
- Your current documents sit alongside three outdated versions and nobody knows which one governs
- Your intranet search only finds what you already know the exact title of
- Your experts spend their time hunting the right paragraph in the right PDF
What it gets you
Versioned
Source of truth
When a document is replaced, the old one leaves the index. No more answers drawn from a withdrawn standard.
Partitioned
Confidentiality
Restricted documents only surface for people who already have access to them in your systems.
What is included
- An audit of your sources: formats, volumes, duplicates, competing versions
- An automated ingestion pipeline from SharePoint, Drive, a DMS or a network folder
- Semantic chunking and vector indexing tuned to your document types
- Hybrid search — meaning and exact terms — so a product reference is never missed
- A consultation interface and an API to plug the base into your other tools
- A regression test set to verify that right answers stay right
How it runs
01
Document audit
1 to 2 weeksWe measure what you actually hold: volumes, formats, duplicates and contradictory documents. This is often the project’s first surprise.
02
Ingestion pipeline
2 to 4 weeksAutomating retrieval, chunking and indexing, including how updates and deletions are handled.
03
Tuning relevance
2 to 3 weeksWe assemble a set of real questions with your experts and tune retrieval until the answers hold up.
04
Operations
OngoingMonitoring document freshness, tracking poorly served questions and reindexing at the pace of your changes.
Indicative timelines, heavily dependent on the initial state of your documents. Clean documentation halves the project.
The questions you are asking
How is this different from a search engine?
A classic engine hands you a list of documents to open. RAG reads the relevant passages and writes the answer, stating where it came from. You move from "here are twenty PDFs" to "here is the answer, page 34 of the 2026 manual".
Do our documents go to an AI provider?
The index stays with you or on infrastructure you choose. Depending on sensitivity, we configure a model hosted on a private cloud or on your own servers, with no transit through a public service.
What happens when a document is updated?
The ingestion pipeline detects the change, reindexes the new version and removes the old one. That is precisely what separates a maintained knowledge base from a frozen demo.
Can this base feed an assistant or a voice agent?
Yes, and that is the most common setup. The base becomes the shared source of truth: the internal assistant, the public site and the phone agent then all give the same answer.
Your answers already exist. Finding them is the hard part.
We start with an audit of your documents, then build a base tested against your real questions.
The other AI modules