RAG and Knowledge Search with Answers You Can Check

The system searches your documents for passages relevant to a question, then asks a language model to answer using only those passages and to cite them. The model supplies the wording, not the facts. Answers can therefore be checked, and kept current by updating the documents rather than retraining anything.

We carry the permissions from each source system — SharePoint, Google Drive, Confluence or your own application — into the index, and filter every search by the signed-in user’s access before anything reaches the model. We test it with accounts at different access levels, and re-sync permissions as often as the documents themselves.

The assistant says so. If retrieval finds nothing that clearly supports an answer, it replies that it could not find one, shows what it searched, and points to the right team. Those unanswered questions are logged, because they show exactly where your documentation has gaps.

With a test set of real questions and agreed answers, written with the people who know the subject. We score whether the right passages were retrieved, whether each answer is supported by its citations, and whether the system correctly refuses questions it cannot answer. Every change is re-scored before release, and live feedback adds new questions over time.

Scheduled or event-driven syncs pick up new, edited and deleted documents, so a withdrawn policy stops being quoted soon after it is removed. How fresh each source needs to be is agreed up front — a price list may need hourly updates, a handbook daily — and every answer can show when its source was last indexed.

That is common, and better found early. We rank sources by owner, status and date, exclude drafts and archives where you want them excluded, and surface contradictions rather than blending them into one answer. Where two current documents disagree, the fix belongs with the document owner — the system shows them where to look.

Yes, with the right parsing. Scans go through OCR, tables are extracted with their rows and headers intact, and spreadsheets are indexed in a way that keeps each value with its labels. Numeric questions over large tables are often better answered by querying the data directly, and we build that route in where it applies.

In your own cloud account or database, in a UK or EU region if you need one. Passages sent to a model provider go through their business API, whose data terms we confirm for your chosen provider during design. Where documents must not leave your environment at all, we can run open models in your own infrastructure.

A discovery sprint from £2,000 builds the question set, maps your sources and ends with a fixed-price quote for a pilot. Running costs are hosting for the index plus model and embedding fees, which are paid directly to the provider. We estimate the cost per question during discovery, and it usually depends more on answer length than on the number of documents.

It can sit behind one. RAG is the retrieval and answering engine; a chatbot or voice assistant is one way to put it in front of people. An internal search page, a Teams app, a panel inside your CRM and a customer-facing assistant can all share the same retrieval pipeline and the same permissions.