↓ Skip to main content
Yonk-Labs
Yonk-Labs
  1. Projects/

graphrag-demo

·2 mins
Python Postgresql Apache-Age Pgvector Graphrag Rag
Table of Contents

graphrag-demo is a sample app showing what graph-augmented retrieval actually buys you over plain vector search — by running both, plus the combination, against the same question at the same time and showing the timing breakdown for each.

What it demonstrates
#

Every query runs three retrieval strategies in parallel: vector-only (embed the question, cosine similarity via pgvector), graph-only (extract entities, traverse relationships via Apache AGE’s Cypher support), and graph+vector (vector search seeds the traversal, graph expansion pulls in related context, and a re-ranking step combines both signals). Seeing all three answer the same question side by side makes the tradeoff concrete instead of theoretical — vector search is good at semantic similarity, graph traversal is good at structural connections neither approach alone would surface.

Stack
#

PostgreSQL 16 with pgvector (HNSW index) and apache_age (Cypher graph queries) in one instance, a FastAPI orchestrator running the three strategies concurrently, and pluggable LLM (Claude, OpenAI, Ollama) and embedding providers.

Quick start
#

cp .env.example .env
# set ANTHROPIC_API_KEY or OPENAI_API_KEY in .env

docker compose up --build
# open http://localhost:8000

The database seeds itself on first run with roughly 160 documents about a fictional organization, so there’s a working corpus to query immediately.

Blog series
#

The repo accompanies a three-part write-up: why vector search alone falls short, building the graph-aware pipeline, and the head-to-head comparison — all included in the blog/ directory.

Links#