Strata

A Perplexity-style real-time AI search engine that crawls the live web and streams a cited, synthesized answer.

React 19React Router 7Tailwind CSS v4BunExpress 5PrismaPostgreSQLSupabase AuthTavily APIOpenAI API
Source Code →

Problem

Standard chatbots answer from static training data with no live sources, so answers about current events or fast-changing facts are often wrong or unverifiable.

Solution

Built a full-stack AI search app that runs live web search via the Tavily API, streams the synthesized answer token-by-token from OpenAI over Server-Sent Events, and appends verifiable cited sources at the end of the stream.

Impact

Delivers real-time, cited AI answers with sub-second time-to-first-token streaming, secure Supabase OAuth (Google/GitHub), and a type-safe Prisma/PostgreSQL data layer ready for persistent chat history.

Overview

Strata (internally "AI Browzer") is a real-time AI search application: users ask a question, the backend crawls the web live for relevant sources, and the answer streams back with citations attached.

Core Features

  • OAuth Authentication: Google and GitHub sign-in via the Supabase Auth client, with a custom JWT verification middleware protecting every backend endpoint.
  • Live Web Search: Real-time search indexing and retrieval through the Tavily API for up-to-date context on every query.
  • Streamed AI Synthesis: Answers stream chunk-by-chunk from OpenAI's gpt-4o-mini via Server-Sent Events, with cited sources appended at the end of the stream inside a structured <SOURCES> block.
  • Dashboard Workspace: A secured layout showing the authenticated user's profile and session state.

Architecture

The stack runs entirely on Bun as the execution engine for both frontend and backend. Auth flows through Supabase (2-way sync triggers keep the Postgres User table aligned with Supabase's own auth.users), while Express streams the search-and-synthesize pipeline:

  1. Frontend sends the query to POST /strata_ask.
  2. Backend fetches live snippets and URLs from Tavily.
  3. Backend calls OpenAI with the query + search context, stream: true.
  4. Each text chunk is written directly to the response as it arrives.
  5. Once generation completes, the backend appends the source URLs and closes the stream.

The relational schema (Prisma on PostgreSQL) models User → Conversation → Message so full chat history can be persisted and later replayed as context for follow-up questions — the next planned feature.

Planned Features

  • Persisted database chat history and multi-turn, context-aware conversations
  • Per-user daily query credit limits
  • Stripe-based credit top-ups