AI
North AI — RAG-Powered Customer Support Agent
An AI support agent built with LangChain and LangGraph, using retrieval-augmented generation over a knowledge base of thousands of support articles. Deflected roughly 40% of incoming tickets automatically while giving agents a transparent view of every source the model used to answer.
Tags
LangChainLangGraphNext.jsVector Search
The challenge
Support ticket volume was growing faster than the support team, and roughly 60% of tickets were repeat questions already answered somewhere in the client's docs, changelogs, and past tickets. Previous attempts at a simple FAQ chatbot returned confidently wrong answers with no way for agents to verify them.
The solution
- Built a retrieval pipeline that chunks and embeds documentation, changelogs, and resolved tickets into a vector store, with a re-ranking step to surface the most relevant sources for each query.
- Orchestrated the agent with LangGraph so it could decide when to answer directly, when to search for more context, and when to hand off to a human agent instead of guessing.
- Every response streams back with clickable citations to the exact source passages, so support agents and customers can verify the answer instead of trusting it blindly.
- Added a feedback loop where agent corrections automatically flag weak or outdated source documents for the content team to fix.
The results
- The agent now deflects approximately 40% of incoming tickets without any human involvement.
- Average first-response time on tickets that still require a human dropped significantly because agents start with AI-drafted context instead of a blank page.
- Customer satisfaction scores on AI-only resolutions matched human-handled tickets within the first two months.