Grounded discovery across a decade of scattered knowledge

The challenge

Ten years of guides, templates and policy explainers lived across three systems. Staff asked the same questions repeatedly because nobody could find anything, and earlier keyword search failed on natural-language questions.

The solution

A retrieval-augmented discovery layer indexes all structured content with freshness metadata, retrieves by hybrid ranking, constructs grounded prompts under strict separation of instructions and data, and returns every answer with inspectable sources - or an honest refusal when coverage is thin.

Observable outcomes

• ✓Natural-language questions answered with citations drawn only from approved content
• ✓Refusal path prevents confident fabrication when coverage is insufficient
• ✓Source panel exposes exactly which passages informed each answer

Outcomes describe this technical demonstration's behaviour, not client metrics.