A RAG assistant that cut support tickets by a third

Northwind's support team was drowning in repetitive questions. The answers all existed — scattered across a 400-page docs site, a Notion runbook, and three years of resolved Zendesk tickets — but nobody could find them fast enough.

Challenge

The knowledge wasn't missing; it was unsearchable. New hires spent weeks learning where things lived, and customers waited hours for answers that a senior engineer could give in thirty seconds. Off-the-shelf chatbots hallucinated confidently, which for a developer-facing product was worse than no answer at all.

Northwind needed something that would:

• Answer from their content, with citations, and say "I don't know" when it didn't
• Stay current as docs changed daily
• Run inside their own infrastructure for data-residency reasons

What we built

We shipped an internal retrieval-augmented assistant embedded in their existing admin app.

• A Python ingestion pipeline that chunks docs, runbooks, and resolved tickets, embeds them with OpenAI, and stores vectors in pgvector alongside their existing Postgres
• Hybrid retrieval (semantic + keyword) with a re-ranking pass, so exact error strings and conceptual questions both work
• A Next.js chat surface with inline source citations and a one-click "escalate to human" handoff
• Nightly re-indexing wired to their docs CI, so the assistant never drifts from the published source

Every answer links back to the passages it used. When confidence is low, it routes to a human instead of guessing.

"It went from a science project to the first thing our support team opens every morning. The citations are what earned their trust." — VP of Engineering, Northwind SaaS

Results

Numbers below are from Northwind's first quarter on the system and are illustrative of the kind of outcome this work targets.

• −34% inbound support tickets, driven by in-app self-service deflection
• First-response time on escalations down from hours to minutes
• New-hire ramp shortened, since the assistant became the team's institutional memory

The assistant now handles the long tail of "where is this documented" questions, so Northwind's engineers spend their time on the problems that actually need a human.