Service / Knowledge assistants
Answer from
your sources.
A knowledge assistant helps people find and use information in an approved document collection. Retrieval-augmented generation, or RAG, combines a search step with text generation: the system retrieves relevant material, then asks a model to answer using that material. Its usefulness depends on both steps.
1. Define the questions it should answer.
Choose a bounded purpose such as locating an internal procedure or explaining a published product manual. List the questions staff actually ask, the source that answers each question and cases that should be redirected. An assistant intended to find a policy should not quietly become a system that decides exceptions to that policy.
Separate information retrieval from professional judgement. For employment, legal, financial or health-related questions, the assistant can help locate relevant material but should not substitute for qualified advice. Define a route to the responsible team and make that route visible when an answer depends on an individual’s circumstances.
2. Prepare an authoritative collection.
Choose documents with a named owner and a clear publication status. Remove superseded copies from the active collection or mark them in a way the retrieval system can enforce. Retain headings, table context and section references during extraction. A fragment stating an exception may be misleading if its conditions appear on another page.
Chunking divides documents into smaller passages for retrieval. Select passage boundaries that preserve meaning rather than cutting solely by length. Store metadata such as document title, source link, access group and revision status. Scanned PDFs may need optical character recognition, followed by checks where misread characters could change the answer.
3. Match search to the material.
Keyword search is useful for exact identifiers, error codes and product references. Vector search compares numerical representations of meaning, called embeddings, and can help when a question uses different words from the source. Hybrid search combines these approaches. None removes the need to inspect whether retrieved passages actually answer the question.
Re-ranking orders retrieved candidates using an additional relevance assessment. It may improve selection, but adds processing time and another component to maintain. Compare the approaches using your document types and question set. A technically elaborate retrieval chain is not automatically preferable to a simpler search that consistently finds the correct section.
4. Enforce permissions before retrieval.
Apply the user’s access rights when selecting documents, not just when displaying the final answer. If a restricted passage enters the model’s context, it can influence the response even when its citation is hidden. Keep permission metadata synchronised with the source system and test changes in group membership.
Decide how search indexes, embeddings, caches and logs are protected. Derived data may still contain or reveal sensitive information. Deleting a source file should trigger appropriate removal from these related stores. Supplier contracts and data protection arrangements need to cover the actual processing path, including retrieval services rather than only the language model.
5. Make evidence and uncertainty visible.
Provide citations that point to the supporting passage or a usable source location. A citation is not proof that the answer follows from it: evaluate whether the cited content supports each important claim. Avoid combining incompatible instructions from different versions of a document into a single confident answer.
Specify an explicit response for missing or conflicting evidence. The assistant should ask for clarification or say that the collection does not resolve the question. Test questions whose answer is absent, misleading premises and instructions embedded inside documents. Keeping an answer grounded requires behaviour checks as well as an instruction in the prompt.
6. Assign ongoing ownership.
Agree who approves new sources, corrects extraction errors and reviews feedback. Distinguish a search failure from a generation failure: the right passage may never have been retrieved, or the model may have misinterpreted a passage it received. This distinction determines whether to change the collection, retrieval settings or response instructions.
A scoped engagement can include source review, retrieval design, assistant implementation and an evaluation set. Define handover requirements for updating documents and removing access. For background on search mechanisms, the Elasticsearch query documentation provides concrete examples. Select technology only after the information and permission requirements are understood.