Connect assistants to maintained information and test their answers.
Worked examples, decision tables and troubleshooting to help you apply the ideas.
Retrieval helps when answers depend on a changing body of approved knowledge. It supplies relevant material at answer time, but does not automatically make every answer correct.
Read the guide ↗The quality of a knowledge system begins with source authority. A larger collection can make answers worse when it contains conflicting or obsolete documents.
Read the guide ↗Document segments should preserve enough context to answer a question accurately. There is no universal segment size that fits every document type.
Read the guide ↗Metadata narrows the candidate set using known facts such as product, version or access group. It complements semantic matching rather than replacing it.
Read the guide ↗Exact terms and semantic similarity solve different retrieval problems. Product codes, names and specialized phrases may need lexical matching alongside semantic search.
Read the guide ↗If the system retrieves the wrong evidence, polishing the response cannot repair the underlying problem. Diagnose source selection before prompt wording.
Read the guide ↗A citation should lead to the material supporting the nearby claim. Merely listing a document title can create a false impression of verification.
Read the guide ↗An assistant should have a useful response when the knowledge base cannot answer. A clear gap report is better than an invented policy.
Read the guide ↗Retrieval must respect the user's authorization before content reaches the model. Filtering an already-generated answer is too late to prevent disclosure.
Read the guide ↗A knowledge index needs a defined update process. Otherwise, users may receive outdated answers after the source has already been corrected.
Read the guide ↗Deletion needs to propagate through source storage, indexes and relevant caches. A source link disappearing does not prove the content is no longer retrievable.
Read the guide ↗Wrong citations can come from retrieval, context assembly or answer generation. Inspect the intermediate evidence to locate the failure.
Read the guide ↗More context can increase cost and distract from the relevant evidence. Select enough material to support the answer without flooding the model with unrelated text.
Read the guide ↗Reranking can help prioritize the most relevant candidates after initial retrieval. Its value should be tested against the extra cost and latency.
Read the guide ↗A scanned document may need reliable text extraction before retrieval can work. Poor extraction can silently corrupt dates, numbers and tables.
Read the guide ↗Tables encode relationships through rows and columns. Flattening them into loose text can remove the context needed for a correct answer.
Read the guide ↗Conversation history helps maintain context, but it should not automatically become authoritative knowledge. Distinguish user statements from approved sources.
Read the guide ↗A multilingual interface does not guarantee equally good retrieval in every language. Test the complete path from question to source to answer.
Read the guide ↗Conflicting documents should produce an explicit conflict or an approved precedence decision. Hiding disagreement behind a fluent answer creates risk.
Read the guide ↗Production evaluation should include real failures and changing knowledge, not only the original test set. Use feedback to improve both the system and its sources.
Read the guide ↗