The Retrieval Interface: Proactive vs. On-Demand Knowledge Surfacing

Knowledge can be surfaced two ways: on-demand, when a user explicitly asks for it, and proactively, when the system detects that a user is in a context where relevant knowledge exists. Both modes are valuable. They serve different use cases and have different design requirements.

On-demand retrieval is the traditional knowledge base interaction: a user opens the knowledge system, types a query, and reads the results. The user is aware they need information and is actively seeking it. On-demand retrieval design optimizes for query quality, result relevance, and answer completeness. The user is in a pull mode — they came to the system looking for something specific.

Proactive surfacing intercepts the user in their workflow and surfaces relevant knowledge before they have to ask. When a sales rep opens a customer record, the CRM surfaces the relevant knowledge about that account's industry, the products they are evaluating, and the support history. When an engineer starts a pull request that touches a service with known performance characteristics, the code review tool surfaces the runbook for that service. The user did not query the knowledge system — the system detected the context and pushed relevant knowledge into the workflow.

1. Proactive Surfacing Design: Trigger Mapping Identify the workflow steps where relevant knowledge is predictably needed. For a sales workflow: opening a new prospect (surface industry knowledge and competitive positioning), entering a deal stage (surface relevant playbooks and objection handling), scheduling a demo (surface product demo guides for the prospect's use case). For an operations workflow: opening a support ticket (surface similar resolved tickets), starting a deployment (surface the deployment runbook and recent incident history). Trigger mapping is the design step — implement only the triggers where the relevance hit rate will be above 70%.
2. Relevance Threshold Calibration Proactive surfacing with low relevance is friction, not value. If the system surfaces knowledge in nine out of ten contexts where it is not useful, users learn to dismiss the suggestions and eventually stop seeing them. Calibrate the relevance threshold by measuring the click-through rate on proactive suggestions: if more than 60% of surfaced suggestions are dismissed without being read, the relevance threshold is too low. Raise it until the click-through rate exceeds 40%.
3. On-Demand Interface Design For complex queries, the on-demand interface should support conversational follow-up: 'Tell me more about the exception process you mentioned.' The retrieval system maintains context across a conversation session, allowing the user to refine and deepen a query without starting over. Single-turn retrieval is sufficient for simple factual queries. Conversational retrieval is required for complex operational questions where the user needs to explore a topic iteratively.

The proactive surfacing model requires careful design to avoid the noise trap. A system that aggressively surfaces knowledge creates notification fatigue — users learn to ignore all suggestions because too many are irrelevant. The relevance threshold must be tuned based on the specific trigger context. Context from the user's active workflow (the specific record they are viewing, the specific step in the process they are executing) is a much higher-quality signal than context from their general activity. Proactive surfacing calibrated to specific workflow triggers outperforms surfacing calibrated to general activity patterns by a significant margin.