Backend Engine
backend/research_engineLangGraph pipeline: Planner → Executor ⇄ Critic → Synthesizer → Gate → Finalizer. Fully host-agnostic, deterministic checkpointing, citation extraction, and budget guards.
Codebase & Architecture
Source & Inspection
Every component of the Multi-Agent Research Assistant is built from the ground up to be inspectable, verifiable, and self-hostable.
The Verifiability Invariant
The core product claim is verifiability: every citation resolves to a real, verifiable source with a verbatim supporting snippet. Unverified claims are prominently flagged, and eval records are write-once artifacts.
Architecture & Subsystems
The codebase is organized into discrete, deeply decoupled layers:
Backend Application
backend/appFastAPI REST API, WebSocket/SSE streaming, Celery pipeline runner, auth cookies, Postgres/pgvector embeddings, and cryptographic bundle signing.
Frontend Site & App
frontendNext.js 16 + React 19 interface. Multi-target build supporting Standalone Server, Static Documentation Pages, and Tauri Desktop.
Desktop Sidecar
desktopTauri v2 native desktop shell wrapping an embedded Python sidecar with local SQLite checkpoints and local corpus processing.
Evaluation & Benchmarks
backend/evalsRigorous citation-fidelity benchmark harness, baseline comparisons, and support-rate scoring with write-once evidence records.
Running Locally
You can inspect, build, and run the complete research assistant stack on your own machine using standard Docker compose or native development environments.
1. Full Stack with Docker
docker compose -f docker-compose.full.yml up --build
2. Backend Development
cd backend && pytest && ruff check
3. Frontend Development
cd frontend && npm install && npm test && npm run dev