Graph Based RAG for strong reasoning and accuracy across layer

[mprawironego]

Current RAG implementations in Microsoft Fabric (via AI Skills, Azure AI Search integration, and OneLake-backed vector stores) rely predominantly on vector similarity search over chunked documents. While this works well for surface-level semantic retrieval, it has significant limitations for enterprise scenarios that require multi-hop reasoning across structured and unstructured data:

• Loss of relational context: When documents are chunked and embedded, the relationships between entities (customers, products, transactions, regulations, contracts) are lost. Vector search retrieves semantically similar chunks but cannot traverse "Customer A → owns → Account B → linked to → Transaction C → flagged by → Regulation D."

What Would Help

A first-class Graph-Based RAG capability natively integrated into Fabric, including:

1. OneLake Graph artifact — a new item type (alongside Lakehouse, Warehouse, Eventhouse) that stores property graphs in an open format (e.g., extending Delta with graph metadata, or adopting an open standard like Apache AGE or GraphAr), with full Purview lineage and OneLake shortcut support.
2. Automatic graph construction pipelines — Data Factory / Fabric Notebook templates that extract entities and relationships from Silver-layer tables and unstructured documents using Azure OpenAI / Phi models, with human-in-the-loop validation.

Why It Matters

• Accuracy where it counts: Enterprise and regulated customers
• Unlocks Fabric's strategic differentiator: Fabric already unifies data estates in OneLake. Graph-RAG is the reasoning layer that turns that unified estate into a reasoning estate
• Reduces architectural sprawl: Customers currently stitch together Azure AI Search + Cosmos DB Gremlin + custom orchestration