# AI support assistant grounded in product docs

> A RAG-based assistant in the help centre and support inbox that answers from docs and past tickets, and hands off to humans with context.

Source: https://www.mapleorbitsol.com/work/ai-support-assistant

- Service: AI / ML (https://www.mapleorbitsol.com/services/ai-ml.md)
- Client: A B2B software company with a global customer base
- Industry: B2B SaaS
- Region: North America
- Duration: 8 weeks
- Team: 2 AI engineers, 1 full stack engineer

## The challenge

Ticket volume grew faster than the support team, first response times slipped past a day, and most questions were already answered somewhere in the docs.

## How we approached it

1. **Eval set first:** Built 300 real questions with approved answers to measure quality before and after every change.
2. **Retrieval:** Indexed docs, release notes and resolved tickets with hybrid search and re-ranking in pgvector.
3. **Assistant:** Claude-powered answers with citations, confidence thresholds and clean human handoff.
4. **Operate:** Tracing in Langfuse, weekly eval runs and a dashboard for deflection and cost.

## Results

- **42%** of tickets resolved without an agent
- **−65%** first response time
- **93%** answer accuracy on the eval set

## Tech stack

Claude, Python, FastAPI, pgvector, LangChain, Langfuse, Next.js, Sentry