Project 30 / 33

SupportAI

FAQ chat widget with RAG and human handoff

A RAG-based FAQ support assistant that answers from company content and hands off when it cannot.

SupportAI — AI & Automation project cover

04 desktop / 02 mobile

Desktop

SupportAI — desktop interface 01
Desktop 01
SupportAI — desktop interface 02
Desktop 02
SupportAI — desktop interface 03
Desktop 03
SupportAI — desktop interface 04
Desktop 04

Mobile

SupportAI — mobile interface 01
Mobile 01
SupportAI — mobile interface 02
Mobile 02

Overview

SupportAI is a FAQ chat widget using retrieval-augmented generation and human handoff. It answers only from a company's FAQs and admits ignorance through an explicit tool call rather than relying on a similarity threshold alone. Nimbus Analytics is the public product identity used for the demonstration.

Approach

Answer only from a company's FAQs and admit ignorance through an explicit tool call, not merely a similarity threshold.

Highlights

09

  • Gemini gemini-embedding-001 embeddings at 768 dimensions
  • Supabase pgvector with the match_faqs RPC
  • Retrieval threshold 0.5 with the top four matches
  • Groq streaming with Gemini fallback
  • Explicit reportNoAnswer tool
  • Server-authoritative fallback response
  • Two consecutive no-answers trigger handoff and a handoff form
  • 3D Intelligence Core built with Three.js and React Three Fiber
  • Lazy loading, WebGL checks, quality tiers and visibility-aware pausing

Limitations

04

  • Measured performance was 49.8 FPS, below the earlier 60.2 FPS claim.
  • No rate limiting on the public chat endpoint.
  • Concurrent first messages may create duplicate conversations because the session table has no unique constraint.
  • Some visual feedback timing and first-load issues remain.