Warehouse Architecture for GenAI Workloads

GenAI changes warehouse patterns dramatically.

New workload types:

• Embedding generation

• Vector search

• Retrieval-augmented generation (RAG)

• Massive unstructured data indexing

? New Data Types

• Text blobs

• Embeddings (vectors)

• Model outputs

• Prompt logs

? GenAI Architecture Layers

? Raw Corpus Layer

• Documents

• Logs

• Conversations

• Media

Stored in object storage.

? Embedding Layer

Precompute embeddings:

• Store as vector columns

• Partition by creation_date

• Cluster by entity_id

? Retrieval Layer

• Vector similarity search

• Metadata filtering

• Join embeddings with structured features

? Biggest Cost Risk in GenAI

Embedding recomputation.

If you re-embed 50PB repeatedly:

• Explosive compute cost

• Massive shuffle

Best practice:

• Immutable embedding snapshots

• Incremental updates only

? Feature + LLM Integration

Warehouse should provide:

• Clean structured features

• Historical aggregates

• Behavioral signals

LLM consumes curated features — not raw data.