Service · typically 6–8 weeks to a production-grade pilot
Answers from your own knowledge, with the sources attached.
Your company's knowledge sits in SharePoint, Confluence, the DMS, the ERP, tickets, contracts and manuals, and finding it costs hours every day. A knowledge assistant answers from those sources, shows only what each user may see, cites every answer and, when needed, looks things up across systems or takes an action. We build the kind that survives contact with real users.
RAG, explained for decision makers
A language model knows the public internet up to a date. It knows nothing about your contracts, your tickets or your price list. Retrieval changes that.
Retrieval-augmented generation (RAG) is a simple idea. When a user asks a question, the system first searches your documents and systems for the relevant passages, then hands those passages to the language model together with the question. The model writes its answer from that material rather than from memory, and can show where each statement came from.
Done well, this gives you an assistant that is current the moment a document changes, respects who may see what, and can be checked. Done badly, it gives you a chatbot that sounds confident and is wrong just often enough that nobody trusts it, which is worse than no assistant at all.
Agentic RAG goes one step further: the assistant can decide to run several searches, query the ERP for a live figure, compare two documents or draft the email that follows from the answer. That is where an assistant becomes a colleague, and where our agent development work comes in.
- FormatSource analysis, pipeline, assistant, evaluation, rollout
- DurationTypically 6–8 weeks to a production-grade pilot
- ForCIOs, heads of IT, HR, sales, service and knowledge management
- OutcomeA permission-aware assistant with citations, evaluation and monitoring
Where knowledge assistants pay off
Illustrative use cases. The common thread: many people ask similar questions, the answers exist somewhere, and finding them is slow.
Internal helpdesk for IT and HR
Password policies, VPN set-up, parental-leave rules, travel expenses: answered instantly from the current policy, with a link to the source and a ticket created when a person needs to act. See HR and IT.
Sales and tender knowledge
Past proposals, product specifications, reference descriptions and compliance statements at the fingertips of the people writing the next bid. The assistant drafts answers to tender questions from approved sources only. See sales.
Service technicians in the field
Manuals, error codes, the service history of the specific machine and spare-part data, searchable by voice or text on site. Answers cite the manual page so technicians can verify. See manufacturing.
Policy and compliance Q&A
What the procurement policy says about gifts, which retention period applies, which clause governs this case: answered from the current versions with citations, so guidance is consistent and auditable. See legal and compliance.
Onboarding
New colleagues ask the assistant what they would otherwise ask a busy neighbour: how things are done, where things are, who is responsible. Time to productivity drops, and the questions reveal what your documentation is missing.
Contracts and technical documentation
Which contracts contain a change-of-control clause, what the specification says about tolerances, what was promised to this customer: questions across thousands of documents, answered with the passages that support them.
Why document-chat pilots disappoint, and what we do differently
Bad chunking and retrieval
Documents cut into arbitrary pieces lose their tables, headings and context, so the search finds the wrong passages. We parse documents by structure, keep metadata, combine keyword and semantic search and re-rank results before the model sees them.
No permissions
A pilot that indexes everything shows everyone everything, including salary lists and board minutes. Our retrieval checks the user's rights in the source system at query time, so the assistant only sees what the user may see.
No evaluation
Without a test set, quality is a matter of opinion. We build a set of real questions with reference answers, measure retrieval accuracy, answer correctness and citation quality, and run it on every change.
Stale content
An index built once is outdated within weeks. We connect sources through incremental synchronisation, track document versions and show the date of every source in the answer.
Answers without verification
Every answer carries citations that open the source passage, and a clear statement when the sources do not support an answer. Guessing is not a feature.
How we build a knowledge assistant
Source and question analysis
Which questions do people actually ask, and where do the answers live? We collect real questions from tickets, chats and interviews, inventory the sources and check their permission models and quality.
Pipeline and permissions
Connectors for SharePoint, Confluence, your DMS and other sources, document parsing, chunking by structure, indexing in a vector store plus keyword search, and the permission check at query time.
Assistant and interface
Retrieval logic, prompt and model set-up, citations, fallback behaviour and the interface: Teams, Slack, a web app or embedded in the tools your team already uses. Agentic capabilities where the use case needs them.
Evaluation
The question set runs against the assistant; we tune retrieval, chunking and prompts until accuracy and citation quality hold. We test ambiguous, out-of-scope and adversarial questions too.
Rollout and feedback
A pilot group uses the assistant in daily work; feedback and unanswered questions feed the evaluation set and the content owners. Dashboards show usage, answer quality and gaps in the documentation.
Sources, platforms and models we work with
Chosen per project, vendor-neutral. Naming a product means we integrate or evaluate it, not that we are partnered with or certified by its vendor.
Frequently asked questions
Is RAG the same as fine-tuning a model on our data?
No. Fine-tuning changes how a model behaves; it does not reliably teach it facts, cannot be updated as documents change and cannot be restricted per user. Retrieval keeps the knowledge in your systems, current and permission-aware, and lets the model do what it is good at: reading and writing.
How do you make sure users only see what they are allowed to see?
The assistant checks permissions at query time against the source system, for example SharePoint groups or the access lists of your DMS, so retrieval only returns passages the user could open themselves. No separate copy of your permissions that drifts out of date, and dedicated test cases in the evaluation set.
Can the assistant also take actions?
Yes, when the use case calls for it. An agentic assistant can create a ticket, draft an email, look up a live figure in the ERP or compare two contract versions. Actions with consequences wait for confirmation. This is where knowledge assistants and agentic workflows meet.
Where does our data go?
Where you decide. Options range from EU regions of Azure, AWS or Google Cloud to EU-based model providers and on-premises set-ups with open-weight models. Index and logs stay under your control in every option. Our article on sovereign AI and EU hosting compares the trade-offs.
Do we need to clean up our documentation first?
Not entirely. The assistant works with what you have and, usefully, shows where documentation is missing or contradictory. Where sources are chaotic, our data foundations work sets up document pipelines and ownership rules alongside the assistant.
Related services
Custom AI agent development
Single- and multi-agent systems designed and built for production: tool use, memory, MCP servers, evaluation suites, guardrails and EU deployment.
Data foundations for AI
Pragmatic, incremental data work for agents: inventory, document pipelines, APIs, search, definitions and access control, one use case at a time.
AI governance, EU AI Act & GDPR
An AI register, risk classification under the EU AI Act, GDPR-aligned processes and a usage policy your teams will actually follow, built together with your lawyers and your data protection officer.
Sovereign AI: options for EU-hosted language models
US providers with safeguards, hyperscaler EU regions, European providers or self-hosted open-weight models: how to choose by data sensitivity and use case.
Let's find the first workflow worth automating.
A 30-minute intro call, no slides and no obligation. We listen, ask about your processes, and tell you honestly where AI agents would pay off and where they would not.