Module 5: AI/ML & Vector Databases

---

Overview

This module covers AI/ML infrastructure for data platforms, including vector databases for similarity search, LLM Operations (RAG architecture, prompt engineering, embeddings at scale), and feature stores for centralized feature management.

---

Module Contents

Vector Databases

Document	Description	Status
Vector Databases Overview	Comparison and selection guide	✅ Complete
Pinecone Guide	Managed vector database	✅ Complete
Milvus Guide	Open-source vector database	✅ Complete
pgvector Guide	PostgreSQL vector extension	✅ Complete
Weaviate Guide	Knowledge graph vector database	✅ Complete

LLM Ops

Document	Description	Status
LLM Ops Overview	LLM operations overview	✅ Complete
RAG Architecture	Retrieval Augmented Generation	✅ Complete
Prompt Engineering Pipelines	Prompt management and deployment	✅ Complete
Vector Embeddings at Scale	Scalable embedding generation	✅ Complete

Feature Stores

Document	Description	Status
Feature Stores Overview	Feature store comparison	✅ Complete
Feast Guide	Open-source feature store	✅ Complete
Hopsworks Guide	Enterprise feature store platform	✅ Complete

---

RAG Architecture Overview

RAG Patterns:

• Naive RAG: Simple retrieve and generate pipeline
• Advanced RAG: Query rewriting, reranking, context compression
• Hybrid Search: Vector + BM25 for best results
• Metadata Filtering: Pre-filter for better performance

---

Vector Database Comparison

Feature Comparison

Database	Type	Deployment	Index Types	Max Dimensions	Cost	Best For
Pinecone	Managed	Cloud-only	HNSW	20,000	$$	RAG, production, managed
Milvus	Open Source	Self-hosted/Cloud	IVF, HNSW, FLAT	32,768	$	On-premises, K8s
pgvector	Extension	Self-hosted	IVFFlat, HNSW	2,000	$	PostgreSQL shops
Weaviate	Open Source	Self-hosted/Cloud	HNSW	Unlimited	$-$$	Knowledge graphs

Selection Criteria

---

Feature Store Architecture

Feature Store Benefits:

• Consistency: Same features in training and inference
• Reusability: Share features across models
• Version Control: Track feature changes over time
• Point-in-Time Correctness: Avoid data leakage
• Governance: Feature ownership, documentation, lineage

---

Cost Considerations

Vector Database Costs

Factor	Impact	Optimization
Vector dimension	Storage + compute	Use lower dimensions (384 vs. 1536)
Index type	Memory vs. accuracy	HNSW for balance
Deployment	OpEx vs. CapEx	Open source for scale
Region	Data transfer	Colocate with data

Embedding Generation Costs

Model	Cost per 1M tokens	Quality	Speed
OpenAI ada-002	$0.10	Excellent	Medium
OpenAI text-embedding-3-small	$0.02	Very good	Fast
Cohere embed-v3	$0.10	Very good	Fast
Sentence Transformers	Compute cost	Good	Medium

Feature Store Costs

Platform	Cost	Notes
Feast	Free (self-hosted)	Open source
Hopsworks	Usage-based	Managed service
Tecton	Custom	Enterprise
Vertex AI	Usage-based	GCP integration

---

Learning Objectives

After completing this module, you will:

1. Select vector databases: Pinecone vs. Milvus vs. pgvector vs. Weaviate
2. Implement RAG: Retrieval Augmented Generation patterns
3. Scale embeddings: Batch and real-time embedding pipelines
4. Use feature stores: Feast, Hopsworks for ML feature management
5. Optimize AI/ML costs: Model selection, deployment, serving

---

Module Dependencies

---

Quick Start

Vector Databases

1. Start with vector databases overview for comparison
2. Choose based on use case:
   • Production RAG: Pinecone
   • Self-hosted: Milvus
   • PostgreSQL shops: pgvector
   • Knowledge graphs: Weaviate

LLM Ops

1. Learn RAG architecture for LLM systems
2. Implement prompt engineering pipelines for production
3. Scale embeddings with batch processing and caching

Feature Stores

1. Understand feature store patterns
2. Choose platform:
   • Open-source: Feast
   • Enterprise: Hopsworks

---

Next Steps

1. Study Vector Databases
2. Learn RAG Architecture
3. Implement Feature Stores
4. Proceed to Module 6: CI/CD for Data

---

Estimated Time to Complete Module 5: 8-10 hours

Total Files: 12 markdown files with 60+ diagrams