RAG development services: an AI knowledge base built on your own documents
RAG development services turn the documents your company already has, such as manuals, policies, contracts, project files, product data and old tickets, into an AI knowledge base that answers questions in plain language and shows the source of every answer. We build retrieval augmented generation applications that respect the permissions of the systems the documents come from, stay current when documents change, and say "not found" instead of inventing an answer. Hosted in the EU, and your company owns the code.
- You see the result before it goes live
- Fixed price for the agreed scope, in writing before work starts
- Source code, domain and accounts in your company's name
Before and after an AI knowledge base
A service technician needs the torque setting for a part on a machine model that was updated two years ago. Today the answer is somewhere in a manual on a shared drive, a revision note in an email and the head of a colleague who is on holiday. The search takes half an hour, or the technician guesses.
With an internal knowledge assistant, the technician asks the question in the words they would use with a colleague. The system finds the manual, the revision note and the service bulletin, answers from them, and shows links to the three passages, with the newest revision first. If the documents disagree or nothing matches, it says so and names the person responsible for that product. The same pattern works for HR policies, sales proposals, legal templates, IT procedures and customer history.
It is a common starting point for AI consulting work, because the value is visible in the first week and the risks are well understood. If you are weighing retrieval against training a model, read RAG vs fine-tuning.
What we build in a RAG application
- Connectors to your sources. SharePoint and OneDrive, Google Drive, Confluence, file shares, ticketing systems, CRM notes and databases, synchronised on a schedule or on change.
- Document processing. Text and tables read from PDFs, Office files and scans, split into passages that keep headings, versions and dates.
- Search that combines meaning and words. Vector search for meaning and keyword search for part numbers, names and codes, with ranking tuned on your own questions.
- Permission filtering. Each passage carries the access rights of its source document, and results are filtered for the asking person before the model sees them.
- Answers with citations. The model is instructed to answer only from the retrieved passages and to link each one; no source, no answer.
- Interfaces. An intranet search page, an assistant in Teams or Slack, or an interface your own applications use; see API integration.
- Audit log and feedback. Questions, sources used and user ratings recorded, so weak spots in the content become visible.
Permissions and leaks: the security side of RAG
The OWASP Gen AI Security Project lists Vector and Embedding Weaknesses among its top ten risks for language model applications in 2025 (LLM08:2025). It explains that weaknesses in how vectors and embeddings are generated, stored or retrieved can be exploited to inject harmful content, manipulate outputs or access sensitive information, and names unauthorised access and data leakage through inadequate or misaligned access controls, and cross-context leaks in environments shared by several groups of users, as common examples (OWASP, read on 12 November 2026).
Our design answers both. Permissions are not rebuilt by hand; they are read from the source system with every synchronisation, so a person removed from a SharePoint site loses access in the knowledge base too. Departments, clients or subsidiaries that must not see each other's content are kept in separate indexes, not just separate filters. And documents are treated as data: an instruction hidden inside a document cannot change what the assistant is allowed to do.
Personal data and deletion
A knowledge base built from company documents will contain personal data, in emails, HR files and customer notes. The GDPR requires that personal data is kept in a form that permits identification for no longer than necessary (Regulation (EU) 2016/679, Article 5(1)(e)) and gives people the right to obtain the erasure of their personal data without undue delay where one of the grounds applies (Article 17(1)). Both were read on 12 November 2026.
In practice that means the index must never outlive the source. We build deletion into the synchronisation: when a document is deleted or reaches the end of its retention period in the source system, its passages leave the index and stored answers on the next run. Sources that should not be in the knowledge base at all, such as personnel files or health data, are excluded at the connector, and a data protection impact assessment is prepared where the content requires one; see our DPIA guide. Under the EU AI Act, an assistant that people interact with directly must tell them that they are interacting with an AI system unless that is obvious (Regulation (EU) 2024/1689, Article 50(1)), and the European Commission's AI Act page states that the transparency rules apply from August 2026 (read on 12 November 2026).
Accuracy, measured before launch
Retrieval systems fail in two ways: they find the wrong passage, or they find the right one and the model answers badly. We measure both. Your experts write down a set of real questions with the answers and sources they agree on. The test version runs every question and reports, per topic, whether the right source was found and whether the answer matched. Weak topics usually point to content problems, such as outdated or duplicated documents, which we list for the content owners. The same test set runs after every change to the model, the search or the sources, so quality cannot quietly drift.
Ready-made AI search or custom RAG development?
Ready-made products, such as Microsoft 365 Copilot, Glean and the AI search features of Confluence and Notion, work well when your documents live mainly in their own suite and the standard permissions are enough. Custom RAG application development pays off when content is spread across several systems, when part of it lives in databases or line-of-business systems, or when answers must follow your own rules, formats and data location.
| Situation | Ready-made AI search | Custom RAG application |
|---|---|---|
| Content mainly in one suite | Usually the quickest route | Rarely worth building |
| Sources across drives, databases and business systems | Connectors of varying depth | Connectors built per source |
| Separate clients or subsidiaries | Shared tenant settings | Separate indexes by design |
| Answers inside your own applications | Limited | An interface your apps use |
| Model choice and EU data region | The vendor's choice | Your choice, changeable later |
For a customer-facing assistant on your website, the same retrieval layer can power a bot; see chatbot development. For assistants that also take actions, see AI agent development, and for the broader view of language technology, our guide to NLP applications.
How we build it
- Map the content. Sources, owners, permissions, retention rules and the questions people actually ask, written down for your approval.
- Scope and fixed price in writing. Sources, user groups, interfaces and what the first phase leaves out, with a fixed price for that agreed scope.
- A test version on your real data. Two or three sources indexed with their permissions, and the test set of questions measured.
- Pilot and acceptance. One team uses the assistant in daily work, and the phase is accepted against the written scope and the test results.
- Roll out and measure. More sources and teams, with search time, answer ratings and "not found" questions compared with the start.
- Maintenance. Monitoring, model updates, test set reruns and new sources, on a monthly plan you can cancel with one month's notice.
As a standard term, the source code, configuration and test sets sit in a repository in your company's own account, and the index and hosting are in your name in an EU data region. Our AI consultant review helps decide which content to start with. See prices.
Questions we get
What is RAG?
Retrieval augmented generation. Instead of relying on what a language model learned in training, the system first searches your own documents for the passages relevant to a question, then asks the model to answer from those passages only and to show where each part of the answer came from. The model stays general; the knowledge stays yours.
How is an AI knowledge base different from a chatbot?
A chatbot is a conversation channel, usually for customers. An AI knowledge base is the search and answer layer over your own content, usually for employees, and it can sit behind a chatbot, an intranet search box, a Teams or Slack assistant or another application. For a customer-facing bot, see our chatbot development page.
Will employees see documents they should not see?
No, if it is built properly, and this is the first thing we build. The permissions of each document are read from the source system, such as SharePoint, Google Drive or Confluence, and every search is filtered by the asking person's rights before anything reaches the model.
What happens when a document is deleted or changed?
The index follows the source. Changed documents are re-indexed, and deleted documents are removed from the index and from stored answers on the next synchronisation, so the knowledge base does not keep content the source system no longer holds.
How do you make sure the answers are right?
Every answer links the passages it used, so a reader can check it in seconds. Before launch we build a test set of real questions with answers your experts agree on, and measure on it how often the right source was found and the answer was correct. When the sources do not contain the answer, the assistant says so instead of guessing.
Does our data train the AI model?
No. The models run in your account in an EU region, for example through Azure OpenAI, AWS Bedrock or Google Vertex AI, or as an open-weight model on your own infrastructure, under terms where your content is not used for training.
Get a written quote
Tell us in a few lines what you need. We reply within one working day with a written quote: scope, timeline and a fixed price or an hour estimate. Asking is free and commits you to nothing.