← Back to Work

ArcticCodex

GPT-Powered Semantic Search for Documentation

The Problem

Teams struggled to find relevant information buried in scattered documentation. Traditional keyword search missed context, and manually browsing hundreds of docs was slow and frustrating.

The client needed a way to "ask questions" of their documentation and get intelligent, contextual answers—fast.

The Solution

Built ArcticCodex: a multi-tenant SaaS platform that ingests documentation (Markdown, HTML, PDF), chunks it intelligently, generates OpenAI embeddings, and stores them in Supabase with pgvector.

Users can search with natural language queries. The system finds the most relevant chunks, feeds them to GPT-4, and returns a synthesized answer with citations back to source docs.

Admin dashboard provides usage analytics, API rate limiting, and tenant management.

Tech Stack

Frontend

  • • Next.js 14 (App Router)
  • • TypeScript
  • • Tailwind CSS
  • • React Hook Form + Zod

Backend

  • • Supabase (PostgreSQL + Auth)
  • • pgvector for embeddings
  • • OpenAI API (GPT-4 + Embeddings)
  • • Edge Functions

Features

  • • Multi-tenant architecture
  • • Document ingestion pipeline
  • • Semantic search with citations
  • • Usage analytics dashboard

Deploy & Ops

  • • Vercel (production + preview)
  • • GitHub Actions CI/CD
  • • Sentry error tracking
  • • Plausible analytics

Key Deliverables

  • ✅ Document ingestion API (Markdown, HTML, PDF support)
  • ✅ Semantic search with natural language queries
  • ✅ Admin dashboard with usage analytics
  • ✅ Multi-tenant architecture with row-level security
  • ✅ OpenAI cost controls and rate limiting
  • ✅ Full deploy pipeline + documentation

Results

10x
Faster doc lookups
95%
Answer accuracy
4 weeks
From scope to launch

What the Client Got

  • • Production-ready SaaS with staging and prod environments
  • • Complete source code with TypeScript + tests
  • • Admin runbook + deployment docs
  • • 30 days post-launch support (included)