Canonical: https://nexuri.de/en/guides/knowledge-assistant Language: en Practical guide · Company knowledge # An AI knowledge assistant that shows its sources. Procedures, product knowledge and internal documents only help when your team can find them. An internal assistant can search approved content and prepare answers with sources. Data quality, access rights and an honest response when information is missing matter. [Discuss a knowledge assistant](https://nexuri.de/en/ai-check) [Explore the example](https://nexuri.de/en/guides/knowledge-assistant#example) 01 Begin with a defined knowledge area 02 Plan sources and permissions together 03 Evaluate answers with real questions 01 / A worked example ## One question. The relevant source. An illustrative internal assistant workflow. The question and documents are examples, not customer data or proven results. Example internal question > “Which documents do we need when handing over a new property?” - 01 Check access. The signed-in person is identified. Only approved content they are allowed to see can be included in the search. - 02 Find relevant content. Search returns relevant sections of current procedures and checklists. Document version and source location are retained. - 03 Prepare an answer with sources. The assistant summarises the retrieved content and links to the underlying documents. The person can check the answer against the original. - 04 Make gaps clear. If there is no suitable basis or documents conflict, the assistant should make that visible and refer to the responsible person. Sources make answers easier to verify but do not guarantee correctness. Your team needs suitable quality checks and remains responsible for important decisions. 02 / The principle ## RAG combines search with a readable answer. Retrieval-Augmented Generation, or RAG, adds relevant content from a knowledge source to an AI request. For an internal assistant, that means finding the right documents first and grounding the answer in them. In short Retrieval-Augmented Generation (RAG) first finds relevant passages in approved documents, then has the AI ground its answer in them. Training a custom model is usually not required; quality depends mainly on current, readable sources. ### The source is the starting point. Reliable answers need current, readable content. Outdated documents, scans without extracted text and conflicting versions should be reviewed first. ### A defined knowledge area. Start with a collection of approved procedures, for example. Typical questions, missing content and ownership can then be evaluated deliberately. ### Changes are part of operation. New versions, removed documents and permission changes need to be reflected in the search system. That requires an agreed update process. 03 / Trust & access ## An answer must not create a new permission. The assistant should respect your existing access boundaries. Those requirements belong in technical planning, rather than only in a text instruction to the AI. In short A knowledge assistant may only use content the signed-in person is already allowed to see. Permission filtering belongs in the search itself, not in an instruction to the AI. Data flows to model and hosting providers are clarified before launch. ### Check rights before answering. Search and document retrieval must be filtered by the signed-in person’s permissions. Source links and previews must not expose protected information either. ### Clarify data flows. Which content goes to which model or hosting provider? Storage location, retention, agreements and your organisation’s requirements are clarified before use. ### Documents are not instructions. File and message content must not simply change access rules or trigger other actions. A read-only knowledge assistant can deliberately be kept separate from workflows that change records. 04 / Effort & value ## What really influences cost. A sound cost estimate needs a clear scope. One approved knowledge area is different from several systems with different permissions and many document formats. In short The cost of an AI knowledge assistant comes from three areas: one-off document preparation and integration, ongoing operation (model calls, search, hosting) and continuous quality checks and upkeep. A clearly scoped knowledge area keeps all three manageable. ### Preparation & integration. Document quality, scans, metadata, integrations and permission handling determine a substantial part of the setup effort. ### Usage & operation. Model calls, search, storage, hosting and updates require ongoing resources. Request volume and expected answer quality influence the appropriate solution. ### Quality checks & upkeep. Evaluate real team questions: relevant sources, correct statements, handling missing knowledge and access protection. New content and feedback then need a responsible owner. Key terms ## The vocabulary of a knowledge assistant. RAG (Retrieval-Augmented Generation)A technique in which relevant content is first retrieved from a knowledge source and passed to the AI model as the basis for its answer. ChunkingSplitting documents into smaller sections so search can find the relevant passage rather than an entire manual. Permission filterA restriction of search to documents the signed-in person is allowed to see. It is applied before answering, not afterwards. Prompt injectionAn attack in which text inside a document or message tries to override the AI’s instructions. Document content is therefore treated as data, not as commands. For your first conversation ## A useful pilot starts with useful questions. Choose a manageable area with known content and responsibilities. - An approved collection of current documents - Typical team questions, including unanswerable ones - Existing roles and access rules - Someone who can judge sources and answer quality Relevant service ## From an internal assistant to a connected workflow. Our AI & automation service covers internal assistants, RAG, copilots and AI agents. The right solution depends on your team and its specific process. [Explore AI & automation](https://nexuri.de/en/services/ai-automation) Common questions ## Knowledge with clear boundaries. Do we need to train our own AI model? For many knowledge assistants, RAG is a useful starting point: relevant content is retrieved and added to the request. Whether other techniques are needed depends on the task, data and quality requirements. Can the assistant guarantee an answer to every question? No. Answers depend on available content, retrieval and the model. A good pilot explicitly tests questions that cannot be answered. Missing knowledge should be acknowledged rather than replaced by an unsupported answer. Can we start with just a few documents? Yes, a defined knowledge area can make a useful pilot if it covers real team questions. Quality, freshness, appropriate permissions and clear answer evaluation matter more than a large document count. Explore further ## Use knowledge where the work happens. [### Automating customer enquiries Find information, prepare a draft and add the appropriate approval.](https://nexuri.de/en/guides/customer-requests) [### For real estate teams Connect property information, documents and contacts effectively.](https://nexuri.de/en/solutions/real-estate) [### Embeddings & semantic search The search technology behind a knowledge assistant: vectors, hybrid search, reranking.](https://nexuri.de/en/services/semantic-search) [### Private AI Run the assistant with self-hosted models on your own infrastructure.](https://nexuri.de/en/services/private-ai) Technical background from Microsoft Learn - [Microsoft Learn: Retrieval-augmented generation (RAG)](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) - [Microsoft Learn: Document-level access control](https://learn.microsoft.com/en-us/azure/search/search-document-level-access-overview) From knowledge to practical value ## Which question does your team hear again and again? Bring a knowledge area and typical team questions. In a free AI opportunity check, we discuss whether an internal assistant is a useful starting point. [Request a free AI check](https://nexuri.de/en/ai-check) A no-obligation initial conversation. Implementation and scope are agreed separately.