- [NetApp Hybrid RAG (BM25-based) Deployment Guide](./docs/projects/hybrid-rag-bm25/README.md)

- [NetApp Graph RAG Deployment Guide](./docs/projects/graph-rag/README.md)

- [NetApp BM25-based/Document RAG Deployment Guide](./docs/projects/document-rag/README.md)

This project captures a **Document-centric Retrieval-Augmented Generation (RAG)** architecture developed and validated by **NetApp**.

The focus is on building RAG systems that are **explainable, deterministic, and governance-ready** from day one. Instead of defaulting to vector-only retrieval, this architecture uses **BM25 lexical search**, enriched with **explicit entity extraction**, to make every retrieval decision observable and reproducible.

The complete reference implementation, including open source code and step-by-step guides, lives here:

👉 **[https://github.com/NetApp/document-rag-guide](https://github.com/NetApp/document-rag-guide)**

This page serves as the **NetApp-specific overview and entry point**.

Most RAG stacks begin with embeddings and end with uncomfortable questions:

* Why did this document match?

* Which terms mattered?

* Can we reproduce this result tomorrow?

* Can we prove compliance?

Document-based RAG flips that model.


Instead of treating retrieval as an opaque side effect of embeddings, it treats retrieval as a **first-class, auditable system**.

Key reasons this approach works:

* **Explainability by default**: BM25 matches explicit fields and terms. You can point to the exact reason a document was retrieved.

* **Deterministic behavior**: The same query over the same data produces the same result. No hidden ranking drift.

* **Reduced hallucinations**: LLM responses are grounded in retrieved documents, not semantic "near matches."

* **Clear governance boundaries**: Explicit document metadata, entity fields, and retention policies make audits practical instead of theoretical.

Vectors still exist, but only as **augmentation**, never as the sole authority.

NetApp extends Document RAG with **enterprise-grade data management and storage capabilities** that turn a clean design into a deployable system.

Key NetApp-specific enhancements include:

* **Dual-tier memory model**

* **Long-Term (LT)**: authoritative, durable document store

* **HOT (unstable)**: short-lived, user- or session-specific working set

* **Governance-driven isolation**

* HOT exists to enforce retention, policy asymmetry, and blast-radius control

* LT remains stable, conservative, and audit-ready

* **High-performance locality**

* NetApp FlexCache keeps frequently accessed documents close to compute

* Cache eviction is explicit and policy-driven, not accidental

* **Enterprise resilience**

* SnapMirror and MetroCluster support replication and disaster recovery

* Snapshots enable point-in-time audits of "what the AI knew"

* **Safe experimentation**

* FlexClone enables instant copies of indices for testing new analyzers or embeddings without impacting production

The result is a Document RAG architecture that aligns with how enterprises already manage data: **explicit, observable, and controlled**.

This page is intentionally high-level.

For full technical details, code, and deployment guidance, visit the main project:

👉 **[https://github.com/NetApp/document-rag-guide](https://github.com/NetApp/document-rag-guide)**

There you'll find:

* A fully open source, community-runnable implementation

* An enterprise architecture with HOT/LT separation and promotion workflows

* Clear patterns for explainable, compliant retrieval

If your goal is AI you can **explain, reproduce, and defend**, start there.

This project documents a **Graph-based Retrieval-Augmented Generation (Graph RAG)** architecture developed and tested by **NetApp**.

The goal is simple: show how enterprises can build AI systems that are **explainable, governable, and production-ready**, not just clever demos. Instead of relying only on vector embeddings, this architecture uses **knowledge graphs with explicit relationships**, combined with a dual-memory model that separates authoritative knowledge from fast, conversational context.

The full reference implementation, including open source code and detailed walkthroughs, lives here:

👉 **[https://github.com/NetApp/graph-rag-guide](https://github.com/NetApp/graph-rag-guide)**

This repository serves as the **NetApp-focused entry point** and architectural overview.


Traditional RAG pipelines usually start and end with vector search. That works for similarity matching, but it breaks down when teams need:

* Clear explanations for why an answer was returned

* Auditable data lineage and provenance

* Multi-hop reasoning across related facts

* Strong governance and compliance controls

Graph RAG addresses these gaps by storing knowledge as **nodes and relationships** instead of opaque embeddings.

Key advantages include:

* **Reduced hallucinations**: Responses are grounded in explicit graph paths, not nearest-neighbor guesses.

* **Explainability by design**: Every answer can be traced through readable graph queries.

* **Better governance**: Provenance, confidence, and promotion logic live directly in the data model.

* **Multi-step reasoning**: Graphs naturally support traversals across documents, entities, and concepts.

This architecture treats retrieval as a **first-class system**, not a side effect of embeddings.

NetApp extends the core Graph RAG design with **enterprise-grade data and storage capabilities** that make it practical at scale.

Key enhancements include:

* **Dual-memory architecture**

* Long-term memory for authoritative, durable knowledge

* Short-term memory for fast, conversational context

* **High-performance caching**

* NetApp FlexCache enables microsecond-level access to hot graph data

* Cached data expires automatically to prevent stale knowledge

* **Data mobility and resilience**

* SnapMirror provides replication and recovery across sites

* Storage follows workloads, not the other way around

* **Promotion and reinforcement workflows**

* Frequently used or validated facts are promoted from cache to long-term memory

* Confidence, provenance, and audit metadata are preserved end-to-end

* **Operational readiness**

* Designed to integrate with streaming pipelines and production infrastructure

* Supports regulated environments where traceability is non-negotiable

The result is a Graph RAG architecture that aligns with real enterprise constraints: performance, governance, and scale.

This page is only a summary.

For full architecture diagrams, implementation details, and runnable examples, visit the main project:

👉 **[https://github.com/NetApp/graph-rag-guide](https://github.com/NetApp/graph-rag-guide)**

There you'll find:

* A community, open source reference implementation

* An enterprise-grade architecture with promotion and governance patterns

* Clear upgrade paths from laptop demos to production deployments

If you're building AI systems that need to be trusted, explained, and operated long-term, start there.

**BM25 + Vector Retrieval with Governance Built In**

This project highlights a **Hybrid Retrieval-Augmented Generation (Hybrid RAG)** architecture developed and validated by **NetApp**.

The design combines **BM25 lexical search** for deterministic, explainable grounding with **vector embeddings** for semantic coverage. The result is a retrieval system that balances **precision and recall** while remaining **observable, auditable, and enterprise-ready**.

This page provides a NetApp-focused overview of the architecture and its enterprise implications.

The full open source reference implementation lives here:

👉 **[https://github.com/davidvonthenen-com/hybrid-rag-bm25-with-ai-governance](https://github.com/davidvonthenen-com/hybrid-rag-bm25-with-ai-governance)**

Pure vector RAG is good at "semantic vibes" but weak at answering hard questions like:

* Why did this document match?

* Which terms actually mattered?

* Can we reproduce this result next week?

* Can we defend it to auditors?

Hybrid RAG addresses those gaps by **anchoring retrieval in BM25 first**, then using vectors as **supporting context**, not the source of truth.


Key reasons Hybrid RAG works:

* **Deterministic grounding**: BM25 provides explicit, traceable matches against known terms and entities.

* **Semantic coverage without drift**: Vector embeddings expand recall for paraphrases and long-tail phrasing without replacing lexical evidence.

* **Explainability by design**: Every result can be tied back to fields, terms, and highlights rather than opaque similarity scores.

* **Lower hallucination risk**: LLM responses are grounded in retrieved documents with clear provenance before any stylistic refinement.

* **Practical governance**: Retrieval behavior is inspectable and reproducible, which matters in regulated environments.

This approach delivers many of the governance benefits people look to Graph RAG for, **without the operational overhead of graph databases or ontology management**.

NetApp extends Hybrid RAG with **enterprise-grade data management and storage primitives** that make the architecture operational at scale.

Key NetApp contributions include:

* **Dual-tier memory model**

* **Long-Term (LT)**: authoritative, durable knowledge store

* **HOT (unstable)**: short-lived, user- or session-specific working set

* **Governance-first tiering**

* HOT exists for retention control, policy asymmetry, and isolation

* LT remains conservative, stable, and audit-ready

* **High-performance locality**

* NetApp FlexCache keeps frequently accessed shards close to compute

* Eviction is explicit and policy-driven, not accidental

* **Enterprise resilience**

* SnapMirror and MetroCluster support replication and disaster recovery

* Snapshots enable point-in-time audits of "what the AI knew"

* **Safe experimentation**

* FlexClone allows instant, space-efficient copies of indices for testing new analyzers or embedding models without touching production

NetApp's role is not to change how Hybrid RAG works logically, but to **make it reliable, governable, and operable in real enterprise environments**.

This page is intentionally concise.

For full technical details, code, and deployment guidance, visit the open source project:

👉 **[https://github.com/NetApp/hybrid-rag-bm25-with-ai-governance](https://github.com/NetApp/hybrid-rag-bm25-with-ai-governance)**

There you'll find:

* A complete Hybrid RAG reference implementation

* Community and enterprise deployment paths

* Detailed explanations of BM25 grounding, vector augmentation, and HOT/LT promotion workflows

If you're building RAG systems that need to be **accurate, explainable, and defensible**, that repository is the place to start.