Neo4j GraphRAG Studio

An interactive Streamlit application built on top of neo4j-graphrag — turn unstructured documents (PDF/TXT/MD) into a Neo4j knowledge graph, then ask questions against it with GraphRAG.

neo4j-graphrag provides the building blocks (SimpleKGPipeline, VectorCypherRetriever, GraphRAG, schema extraction). This project wraps them into an end-to-end workflow with a UI: dynamic schema inference + human-in-the-loop confirmation, a chat interface, subgraph visualization scoped to the actual answer, and one-click Cypher for verifying what you see against the database directly.

Features

• Dynamic schema inference — instead of hand-writing a graph schema up front, the app reads your document and proposes node/relationship types with an LLM. You review and adjust the proposal (add/remove types) before anything is written to the graph.
• GraphRAG Q&A — combines vector similarity search (find relevant chunks) with 1–2 hop graph traversal (pull in connected entities/relationships), so answers aren't limited to what's in a single retrieved text chunk.
• Answer-scoped subgraph visualization — the graph panel only shows entities/relationships that are actually referenced in the answer, not the entire local neighborhood of the retrieved chunk (which can be large and noisy).
• Cypher verification — every answer comes with a ready-to-paste Cypher query so you can confirm the visualized subgraph against Neo4j Browser directly.
• FREE mode — skip schema guidance entirely and let the LLM extract whatever it finds, for quick exploration.
• Three-way benchmark — ask the same question against the same document and compare, side by side: the current GraphRAG pipeline, plain vector RAG (chunk embeddings only, no graph traversal), and stuffing the whole document into the prompt with no retrieval at all. Shows latency and how many characters were actually sent to the LLM for each approach.

Screenshots

Built from the sample Apple document included in input/.

Schema inference — LLM proposes node/relationship types, you confirm before anything is written:

Resulting knowledge graph (66 nodes / 115 relationships from a single article — founders, executives, products, competitors, acquisitions, legal history):

GraphRAG Q&A — the subgraph panel and Cypher verification snippet are scoped to what's actually in the answer, not the entire retrieved neighborhood:

Drilling into a subgraph (Apple's acquisitions — NeXT, Shazam, Beats Electronics, Q.ai, Pixelmator):

Benchmark tab — same document, same question, three approaches side by side:

Architecture

File	Responsibility
common.py	Neo4j driver + LLM/embedding client construction, shared config
Ingest.py	Document → knowledge graph pipeline (SimpleKGPipeline)
query.py	GraphRAG retrieval, answer-scoped subgraph filtering, Cypher generation
benchmark.py	GraphRAG vs. plain vector RAG vs. full-document-in-prompt, for the benchmark tab
app.py	Streamlit UI (schema tab, chat tab, benchmark tab)

Setup

1. Neo4j

Run Neo4j locally via Docker (APOC plugin included):

docker run -d --name neo4j-graphrag-studio \
  -p 7474:7474 -p 7687:7687 \
  -v neo4j_data:/data \
  -e NEO4J_AUTH=neo4j/password123 \
  -e NEO4J_PLUGINS='["apoc"]' \
  neo4j:latest

2. Python environment

Requires Python 3.12+ and uv.

uv sync

3. Configure LLM access

cp .env.example .env

Fill in .env:

• LLM_API_KEY — required.
• LLM_BASE_URL — leave empty to use the official OpenAI API, or point it at any OpenAI-compatible endpoint (Azure AI Foundry's v1 endpoint, a self-hosted gateway, etc.).
• CHAT_MODEL / EMBEDDING_MODEL / EMBEDDING_DIM — match whatever models your endpoint serves.

4. Run

uv run streamlit run app.py

Sample documents (Tesla, Apple — sourced from Wikipedia, CC BY-SA 4.0) are included under input/ so you can try the schema-inference flow immediately.

CLI usage

Both the ingestion and query steps also work standalone:

uv run python Ingest.py                       # ingest everything in input/ using the default schema
uv run python query.py "who are Tesla's main competitors?"

License

MIT — see LICENSE. Sample data under input/ is sourced from Wikipedia (CC BY-SA 4.0).

neo4j-graphrag, neo4j (driver), and streamlit are Apache-2.0. streamlit-agraph is MIT. All are used as ordinary pip dependencies, not vendored.