Your knowledge. In answers you can check.

An AI knowledge base for a specific team and its work. We start with sources and permissions, then test answer quality before rollout.

Illustrative example from technical support

How long is the warranty on spare parts?

24 months from delivery for spare parts ordered under the framework agreement.

Framework agreement · Section 8, paragraph 2
Spare parts supplied under this framework agreement carry a 24-month warranty from delivery.

Your team has the knowledge. Finding it takes work.

Information sits across different tools, document versions and layers of access.

  • Knowledge lives across systems

    Policies, product information and project documents sit in SharePoint, Confluence, file systems, tickets and specialist applications.

  • The same thing goes by different names

    Teams use different terms. Finding the right document can depend on knowing the exact wording.

  • Several versions compete

    When old and new documents sit side by side, it is hard to tell which one applies.

  • Permissions have grown over time

    Review existing access rights before carrying them into a new search system.

We start with four questions. Who asks what? Which sources are authoritative? What can each person access? What makes an answer useful?

Start with a specific knowledge task.

Define who needs an answer, what they need to know and where the answer should come from.

  • Support needs answers with sources

    Teams need information from product documentation, previous tickets and internal guidelines.

  • Engineers search across repositories

    Specifications, decisions and lessons from earlier projects are spread across systems.

  • New hires need controlled access

    Give new team members access to approved onboarding material and process guidance.

  • A prototype is ready for a production review

    The RAG demo works. Permissions, source versions, evaluation and ongoing operation still need to be set up.

Start by documenting knowledge and assigning source owners. Keep qualified people responsible for answers with legal consequences.

From real projects.

For Silver Investment Partners, we built a platform for confidential deal documents, with controlled access and findings linked to their sources. The partners make the investment decisions. Read the Silver Investment Partners case study

At Carano, a support chatbot was one part of a broader AI rollout that included training for 70 employees and a data protection strategy. Read the Carano case study

For more examples and pilot considerations, read our guide to RAG for mid-sized businesses.

Five steps. From real questions to rollout.

  1. Step 1

    Start with real questions.

    We gather typical and difficult questions from your team's day-to-day work, including questions the system should decline to answer.

    Result. A defined knowledge task and a baseline.

  2. Step 2

    Agree sources and permissions.

    We identify an owner, update process and access rules for each source. User permissions apply at retrieval time.

    Result. A source and permissions model.

  3. Step 3

    Design retrieval and answers.

    We test keyword, semantic and hybrid search on your question set. The model receives only the selected content the user can access.

    Result. A retrieval-augmented generation (RAG) architecture for your use case.

  4. Step 4

    Test the search. Then test the answer.

    A fluent answer can hide poor retrieval. We check the passages first, then check what the model says.

    Result. An evaluation set that covers both quality and security.

  5. Step 5

    Prepare the pilot and rollout.

    A limited group of users tries real tasks. Before expanding, we check source updates, monitoring, permissions changes and regression tests.

    Result. A pilot and an operations handover package.

Cite the passage. Respect access.

A useful citation points to the exact passage the user can access. Permission filters run before content reaches the model. When the source material cannot support an answer, the system identifies the gap.
  • Check permissions at retrieval

    The search uses the requester's document permissions. Restricted passages stay out of the model's context.

  • Show the source passage

    Each answer points to the passage and version it uses so people can check the wording.

  • Decline when the evidence is missing

    If no approved passage supports an answer, the system says so and points to the responsible person.

Good answers start with good sources.

Maintain the sources, track errors and test the system again after changes.

  • Quality improves through source maintenance, feedback and retesting.
  • Source owners are responsible for keeping documents current.
  • GDPR obligations depend on how the system is built and used.
  • Measure any time savings against your baseline.
  • Keep experts responsible for decisions where a wrong answer has consequences.

Questions about AI knowledge bases.

Which knowledge task should go first? Let's find out.