Maestro — Graph Rag Orchestration Engine (fastapi + React + Pgvector)

Hi HN, I’m Minwoo, a solo developer from Seoul. Over the past two months I built Maestro, a deterministic orchestration engine that tries to replicate human judgment — not through LLM reasoning loops, but through a Graph RAG that stores and replays how decisions are made.

What it does

Maestro turns every decision (persona, campaign, trend, draft, publication, comment) into a graph node with embeddings in PostgreSQL pgvector. Each connection (edge) represents a relationship — “produces,” “comment_on,” “related_trend,” etc.

When a user asks,

“Show me drafts related to Trend X,” Maestro: 1. Embeds the query using multilingual-e5-base 2. Retrieves the nearest Trend node via vector similarity 3. Expands along relevant edges (related_trend → draft → publication) 4. Surfaces contextual summaries and KPIs — no extra LLM calls required.

The result is a reasoning-aware search engine that rebuilds why something happened, not just what.

How it works

Backend: FastAPI + Celery + PostgreSQL (pgvector) + Redis + SeaweedFS

Frontend: React 19 + Vite + Zustand + shadcn/ui

AI layer: Hugging Face multilingual-e5-base embeddings

Graph RAG: 9 node types, 7 edge types, deterministic traversal

Architecture: DAG executor (idempotent flows) + self-generative adapters

Every action is an idempotent operator. Flows are chained deterministically through a DSL → DAG pipeline. When two flows’ input/output types match, the system connects them automatically — new features emerge with zero code growth.

Example: comments.search + template.create → "Search comments and create reaction template" No new logic written, just an adapter between existing flows.