Customer Experience & Conversational AI Practical insights
Choose which knowledge your chatbot may answer from
Connect a knowledge center to your chatbot with approved sources, conflict rules and answer-level evidence, then test retrieval before expanding coverage.
Connecting every website page to a chatbot can create competing answers. An old launch post, a comparison article and a current product guide may describe different conditions. Build an approved answer collection around the decisions customers need to make, while keeping useful editorial content available for further reading.
Give each source a role
Mark current policies and product documentation as authoritative for their defined scope. Label opinion articles, historical announcements and examples separately. Store language, audience, product version, effective date and content owner where relevant. These labels help retrieval select the right material; they do not repair inaccurate source content.
Resolve contradictions before generation
Suppose a launch article advertises a feature on every plan while current documentation limits it to two plans. Decide which source governs and correct misleading public copy where appropriate. Do not ask the model to average the statements. If two current policy owners disagree, pause that answer path until the conflict is resolved.
Return evidence with useful context
A retrieval result should include the relevant passage, its source and the conditions needed to interpret it. Anthropic’s tool guidance emphasizes meaningful information returned to agents. [1] Apply that principle to answer evidence: include a plan limitation alongside the feature description, rather than returning the headline alone. Link customers to a source they can actually access.
Test retrieval separately from writing
Use questions with known supporting pages, including Dutch phrasing, abbreviations and outdated terminology. First check whether the correct evidence is retrieved. Then check whether the answer preserves its qualifications. Track missing evidence, wrong versions and unsupported conclusions as different defects. This makes fixes more precise than repeatedly rewriting the chatbot’s general prompt. Expand the collection only after an owner can maintain the added material.
Sources and evidence
Sources checked on 4 October 2026. Proposed workflows and hypothetical examples are editorial analysis.