Company knowledge.
Accessible with RAG.
Answers with traceable sources: DAXS connects documents, search and language models into a knowledge solution that accounts for current information and existing access rights.
Answers should lead you to your authorised sources.
An instruction is stored in document management, a technical note in internal documentation and the latest addition in a ticket. Suitable knowledge search can connect these sources and assemble understandable answers. DAXS plans and integrates these solutions, including data import, search, user permissions and ongoing maintenance.
Retrieval-Augmented Generation, or RAG, combines search with a language model. Search retrieves relevant content; the model receives selected passages as context and formulates an answer. In a typical RAG workflow, model weights are not trained with every new document. The knowledge collection and answer model remain separate components.
Internal documentation
Find authorised manuals and operating instructions more easily, summarise a relevant passage and open its source.
Support knowledge
Find established solutions for an enquiry and prepare a verifiable draft reply. Customer and tenant boundaries remain part of retrieval.
Workflows
Bring together information from suitable project documents or policies for a specific task. Business approval depends on the use case.
Import and retrieval are two maintained data flows.
During import, content, source references and permissions are prepared. For a question, the service establishes the user’s identity and searches the permitted collection. Only relevant, authorised results are assembled for the answer.
The architecture can use local or external models. Permitted data flows for embeddings, reranking and answer generation are assessed separately. A locally stored search index does not establish that every subsequent processing step also runs locally.
Good search starts before the first prompt.
Document import involves more than reading a filename. We review supported formats, text recognition, tables and a traceable document structure. Unnecessary duplicates or old drafts should not compete unnoticed with a current operating instruction.
| Step | Task | Key consideration |
|---|---|---|
| Import and preparation | Connect suitable sources, extract content and prepare it in a structured form. | Retain source IDs, versions, timestamps and owners; make errors and unreadable documents visible. |
| Chunking | Divide documents into useful sections. | Respect headings, tables and their business context. Each section needs a reference to its source and permissions. |
| Embeddings and index | Represent content for semantic search and store it with search metadata. | Version the embedding model and index together. Vectors and derived data may also need protection. |
| Hybrid search | Combine semantic similarity with suitable full-text or keyword search. | Exact product numbers, technical terms and questions phrased by meaning have different retrieval needs. |
| Selection and reranking | Order and limit authorised results for the specific question. | Pass only authorised information to an additional reranker; account for the service and approved data flow. |
| Answer and sources | Provide bounded context to the model, answer clearly and link the sources. | The source passage, version and user access must match the result. A citation alone does not prove that every statement was correctly derived from it. |
Knowledge search must respect existing permissions.
An employee without access to a document must not receive its confidential content through a summary either. User and group context therefore come from trusted authentication. The application must not rely on a username supplied by the model or an unrestricted search filter sent with the request.
Depending on the source system and search platform, controls can include inherited ACLs, server-enforced filters or additional object-level checks. Separate index structures may suit strictly segregated collections. What matters is the complete chain: import, permission mapping, retrieval, sharing results, caches and the source actually opened.
A good prompt cannot replace missing access controls.
If identity is unclear, permission information is missing or approvals are no longer valid, the data flow must close according to the agreed protection design. Document text reaches a reranker and model context only after successful checks. Result lists, filenames and answer caches must also avoid disclosing another user’s information.
We explicitly test separate user roles, revoked group rights and other tenants. A document source may also contain manipulated instructions. These remain source content and do not grant tool permissions. If knowledge search should also initiate actions, it receives a separately designed controlled interface.
New content, deletions and permission changes belong together.
A folder does not remain current just because it was imported once. We define how changes are detected, failures retried and obsolete entries removed. This covers new versions, deleted files, revoked permissions, changed sharing links and employees leaving the organisation.
- Synchronisation: suitable connectors, events or scheduled reconciliation with a traceable processing status.
- Source references: stable IDs and versions so references do not accidentally point to a different document version.
- Deletion design: account for affected index sections, embeddings and caches; agree backup retention rules separately.
- Operations: detect failed imports, incorrect permission mapping and unexpectedly old data.
Handling of an outdated or unavailable source system is also agreed. Depending on protection requirements, a warning, a limited answer or temporarily suspending the affected collection may be more appropriate than an answer that only appears current.
Assess retrieval quality and answer quality separately.
A fluent answer may rely on an incorrect or incomplete source passage. We therefore check whether retrieval finds the necessary information, whether the authorised context is sufficient and whether the answer represents that information accurately.
Known question
An authorised question with an established source. Check the retrieved passage, correct version and appropriate answer together.
No suitable source
The system should make missing information clear and offer a useful next step, rather than inventing an apparently supported answer.
Prohibited or outdated source
A user with different permissions and a withdrawn version. Neither document content nor unauthorised hints may appear through an answer or cache.
A suitable test set also includes contradictions, multilingual terms, tables and unusual document formats. Changes to chunking, embeddings, retrieval strategy or the model are tested against it again. RAG can improve access to information; incomplete sources and incorrect answers remain possible limitations.
Start with a suitable knowledge area.
DAXS reviews data sources, permissions and typical questions, develops the import and retrieval flows and connects a suitable application. A limited selection of well-maintained documents with clear owners is often more useful for a pilot than importing the entire file collection at once.
The result is a traceable design with approved data flows, a verifiable question set and documented responsibilities for synchronisation, maintenance and operations. Whether inference runs locally or externally is decided in the infrastructure comparison process. The workplace introduction covers users, training and usage rules.
A defined document collection, responsible contacts, typical questions and the existing permission design. Confidential files are provided only through an agreed, protected channel.
We clarify this in our conversation.
Is RAG the same as a trained company model?
No. RAG retrieves document passages relevant to a question and supplies them as model context. This does not automatically change model weights. Retrieval quality, sources and the answer model need to be assessed together.
How are confidential documents protected?
Identity and permitted document access must be enforced during retrieval and before context is shared. Permissions also apply to the sources displayed. The review includes indexes, caches, exports and logs.
What happens to changed or deleted documents?
The solution needs defined reconciliation with its sources. Changes, permission updates and deletions must reach the search index and affected caches. Obsolete or no longer authorised content must stop being used as current context.
Do citations prevent incorrect answers?
They make checking easier but do not guarantee a correct statement. An answer can misunderstand a passage or summarise it inappropriately. We therefore test source coverage, citation accuracy, contradictions and handling of missing information.
Can a small document collection serve as a pilot?
Yes. A clearly scoped, maintained collection with real questions and known permissions often works better than an immediate full import. The results establish which sources, search methods and operating processes are needed next.
Titles, excerpts and caches need permissions too.
Permissions must be effective before information reaches the model. The review covers result lists, previews, citations, access to originals and answer caches. Group changes and revoked shares are traced through negative tests. An approved search index must not become shared, unrestricted access to confidential sources. The dedicated security page explains oversharing, incorrectly granted rights and current Microsoft controls.
Let's discuss your use case.
What should become easier, which data is involved and which systems are already in place? We use this to develop a practical starting point and a manageable next step.