Welcome to the Graph RAG Guide... a dual‑memory, knowledge‑graph approach to Retrieval‑Augmented Generation (RAG) that delivers answers that are faster, clearer, and governed by design. The repo houses two fully working paths:
- Community / Open‑Source Version - a self‑contained demo you can run on a laptop.
- Enterprise Version - a production‑grade variant that layers in Kafka CDC pipelines, NetApp FlexCache, SnapMirror, and other operational muscle.
By storing knowledge as nodes & relationships instead of opaque vectors, the agent gains traceability, reduces hallucinations, and meets demanding audit requirements.
This project was developed to address significant limitations inherent in traditional vector-based RAG solutions. Specifically:
- Performance Shortfalls: Traditional vector-based RAG, although effective at handling vast unstructured data, often suffers from performance bottlenecks and reduced efficiency during multi-step reasoning tasks. Graph-based RAG, complemented by optimized caching and storage layers, bridges this gap.
- Transparency and Explainability: Vector embeddings are inherently opaque, limiting interpretability and traceability. Graph-based RAG explicitly stores relationships between entities, significantly improving transparency, governance, and auditability.
Key objectives include:
- Provide a reference architecture for graph‑based RAG with explicit short‑term (cache) and long‑term (authoritative) memory.
- Demonstrate how reinforcement‑learning signals can promote proven facts from cache to durable storage.
- Show upgrade paths... from a minimal Python/RAM‑disk demo to an enterprise pipeline with exactly‑once Kafka connectors and enterprise storage.
## Benefits Over Vector-Based RAG
Adopting graph-based RAG addresses several key limitations of traditional RAG Agents:
- Multi-Hop Reasoning: Graph databases naturally support complex queries involving multiple related entities, enhancing retrieval accuracy.
- Bias Mitigation: Explicitly structured relationships simplify detection and correction of biased data.
- Improved Compliance and Governance: Detailed provenance and transparent data relationships facilitate auditability and regulatory compliance.
- Risk Management: Significantly reduced hallucinations due to verified and structured context.
| Capability | Community Edition | Enterprise Edition |
|---|---|---|
| Data movement | Python helper scripts copy sub‑graphs on demand | Kafka Connect Source + Sink connectors stream CDC events between Neo4j tiers |
| Cache backing store | tmpfs / RAM‑disk on the dev box |
NetApp FlexCache for NVMe‑speed reads; resiliency via SnapMirror replicas |
| Operational guarantees | At‑least‑once copies; manual scripts | Exactly‑once semantics, replay, and audit via Kafka offsets |
| Governance hooks | Basic provenance tags | Promotion events logged, schema‑versioned, and instantly traceable |
| Performance | Millisecond‑level on warm RAM cache | Sub‑50 ms at global scale; hot blocks auto‑evicted by FlexCache |
TL;DR: start with the community guide for laptops; switch to the enterprise path when you need 24×7, multi‑site, or compliance.
## Where To Dive Deeper
| Document | What it covers |
|---|---|
| Knowledge Graphs for Better AI Governance | Vision & business case for graph‑based RAG link |
| Community Version Guide | Step‑by‑step setup for the open‑source flavour link |
| Community README | Hands‑on commands & scripts link |
| Enterprise Version Guide | Deep dive on Kafka CDC, FlexCache, SnapMirror link |
| Enterprise README | Production deployment notes link |
| Benchmark Discussion | Discussion on benchmarks for Community Version link |
# 1. Clone the repo
$ git clone https://github.com/<your‑org>/graph‑rag‑guide.git
# 2. Pick your path
$ cd graph‑rag‑guide/community_version # laptop demo
# or
$ cd graph‑rag‑guide/enterprise_version # prod‑ready setup
# 3. Follow the README in that folderQuestions? Open an issue or start a discussion... contributions are welcome!