Lore

Multi-tenant RAG copilot for B2B SaaS support teams. Operators upload docs, customers get instant cited answers via an embeddable chat widget.

Live demo → lore.massimilianoangelone.com

One-click Enter demo — no signup. Two pre-seeded workspaces (Stratos, Nimbus Labs) reset on each session.

What it is

A dashboard for support operators with three jobs: ingest a knowledge base (PDF or Markdown), serve cited answers through an embeddable widget, and track usage across tenants. Every query is scoped to an organisation via Supabase Row Level Security — no application-level tenant code touches data access.

Features

• Document ingestion — PDF/Markdown upload → chunk (512 tok) → text-embedding-3-small → pgvector
• Embeddable widget — <script> tag + API key auth, SSE streaming with source citations
• Multi-tenant by design — RLS at the database boundary, every query org-scoped
• API key management — SHA-256 hashed at rest, plaintext never persisted
• Onboarding flow — first-run workspace setup, document upload guide, API key issue
• Analytics dashboard — token costs over time (Recharts), top questions, avg response time
• Mock mode — three env flags run the app fully offline, no API keys required

Stack

Framework    Next.js 16 (App Router) + React 19 + TypeScript strict
Database     Supabase (Postgres + Auth + RLS) + pgvector
AI           OpenAI gpt-4o-mini + text-embedding-3-small
UI           Tailwind CSS v4 + Base UI + framer-motion + Three.js (login background)
Charts       Recharts
PDF parsing  unpdf
i18n         next-intl (en / it / es)
Tests        Vitest (unit) + Playwright (e2e)
Hosting      Vercel

Architecture

flowchart LR
  A[Operator] -->|upload PDF or MD| B[Ingest]
  B --> C[Chunk 512 tok]
  C --> D[Embed]
  D --> E[(pgvector)]
  F[Visitor] -->|widget| G[/api/chat/]
  G --> H[Embed query]
  H --> E
  E -->|top-k chunks| I[gpt-4o-mini]
  I -->|SSE + citations| F

Loading

Each operator workspace is an org row. RLS policies on documents, chunks, conversations, messages enforce org_id = auth.jwt() ->> 'org_id'. The widget path bypasses session auth and uses API key authentication with SHA-256-hashed tokens.

Run locally

git clone https://github.com/maxange-developer/lore.git
cd lore
pnpm install
cp .env.example .env.local

Zero-config offline mode — flip these four flags in .env.local and skip Supabase + OpenAI setup:

USE_MOCK_AUTH=true
USE_MOCK_AI=true
USE_MOCK_DATA=true
NEXT_PUBLIC_USE_MOCK=true

pnpm demo:seed   # seeds 2 orgs, 5 docs, 5 conversations, 12 messages, 2 API keys
pnpm dev         # http://localhost:3000 → click "Enter demo"

For a real-flavour run with live OpenAI calls, set the four flags to false and provide NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY, OPENAI_API_KEY, NEXT_PUBLIC_APP_URL. Migrations live under supabase/migrations/ and run via the Supabase SQL editor.

Project structure

src/
  app/
    (auth)/            Auth route group (login, signup)
    actions/           Server Actions
    api/               Route handlers (chat SSE, ingest, widget)
    app/               Authenticated dashboard
    onboarding/        First-run setup flow
    widget/            Embeddable widget standalone route
  components/
    chat/              Chat UI + streaming primitives
    conversations/     History views
    dashboard/         Analytics, top questions, costs
    documents/         Upload + management
    embed/             Embeddable widget components
    ui/                Base UI primitives
  lib/
    ai/                Retrieve, chat, embeddings (OpenAI + pgvector)
    auth/              Supabase auth helpers
    db/                Server queries
    demo-state/        Per-session demo reset
    mock/              Offline fixtures mirroring the DB
    supabase/          Client + service-role factories
    validations/       Zod schemas
scripts/               seed-mock, translate-i18n
supabase/migrations/   Database schema
tests/                 Vitest + Playwright

Scope

Lore was built over three weeks as a portfolio engagement to demonstrate a multi-tenant RAG architecture end-to-end. The live deploy serves mock data (two seeded orgs, no real customers) and a demo bypass on login. No paying tenants. The performance and cost figures cited in the case study are operational targets validated against the seeded dataset, not field metrics.

License

MIT.

Author

Built by Massimiliano Angelone — AI-Enhanced MVP Developer, Tenerife.