Self-correcting multi-agent research console. Plan queries, retrieve in parallel, grade the evidence, loop until the critic passes, then synthesize a Markdown brief.
https://veritas-engine-woad.vercel.app/
Status: The public site defaults to Demo mode (simulated). Keys alone do not enable spend — set
LIVE_MODE=true(optionalPUBLIC_RUN_TOKEN/x-run-token). There is no login.
This is the only public repo for the project.
| Console | Cycle |
|---|---|
 |
 |
| Brief | Quality audit |
|---|---|
 |
 |
- LangGraph.js cycle: Planner → Retrieval → Critic → (loop or) Synthesizer
- Critic scores 1–10 and reroutes below 8 until max iteration depth
- Parallel Tavily searches on Live; mock hits on Demo
- SSE stream of node updates into the logstream and graph visualizer
- Custom zero-dependency Markdown renderer (headings, tables, lists, code)
- Copy brief or export
.md - Dark / light console; Demo / Live toggle
| Layer | Technology |
|---|---|
| Frontend | Next.js 14 App Router, React 18, TypeScript, Tailwind 3 |
| Agents | LangGraph.js + LangChain.js (Annotation.Root) |
| LLM (Live) | Gemini 2.5 Flash via @langchain/google-genai |
| Search (Live) | Tavily Search API |
| Streaming | Server-Sent Events from POST /api/research |
| Data on Vercel | Demo repository (simulated retrieval + canned brief) |
| Auth | None |
| Hosting | Vercel |
| CI | GitHub Actions — Vitest, tsc, Playwright |
Start → Planner → Retrieval (parallel) → Critic
↑ │
└── score < 8 ───────┤
▼
Synthesizer → Markdown UI
git clone https://github.com/devtechedge/veritas-engine.git
cd veritas-engine
npm install
cp .env.example .env.local
npm run devOpen http://localhost:3000. Demo mode runs without keys.
npm test
npm run typecheck
npx playwright install chromium
npm run test:e2ePortfolio demo: no login. Demo is forced unless LIVE_MODE=true (and optional run token). Headers, origin checks, rate limits: see SECURITY.md. Live keys stay on the server.
Details: SECURITY.md.
MIT. See LICENSE.