Vector DBs and Semantic Search

Effective vector search depends on more than wiring a vector store into a pipeline—it requires understanding why certain queries succeed, why others fail, and which components are responsible. This lab addresses the gap between using a vector store and tuning one, teaching how embedding models, distance metrics, and chunking strategies each shape retrieval results. With retrieval now central to a wide range of AI applications, these are the patterns you'll use to diagnose retrieval problems and improve quality on your own data.

By the end of this lab, you will be able to:

• Test how embedding models handle difficult language—negation, sarcasm, idioms, and domain jargon—to identify the failure modes most likely to degrade retrieval quality
• Compare how distance metrics, such as dot product and cosine similarity, change what a vector store returns for the same query and dataset
• Evaluate how the choice of embedding model affects retrieval behavior, including how model size influences results
• Apply different chunking configurations to documents and measure their impact on retrieval performance
• Build a repeatable evaluation framework using metrics like recall@k and MRR that can be adapted to diagnose retrieval problems and tune pipelines on your own data

This lab will make use of:

• Python
• Jupyter Notebooks
• Sentence Transformers, FAISS, and associated libraries
• A Windows-based VM