Similarity metrics and thresholds
Cosine similarity is the usual default, but the number it gives you is only meaningful relative to your own corpus. Calibrate the threshold, do not guess it.
- 8m
- Advanced
Video by Rohan-Paul-AI · YouTube
Check your understanding
3 questions. Unlocks when you finish the video.
Overview
What this lesson covers
- Compare cosine, dot product, and euclidean distance
- Calibrate a relevance threshold on real queries
- Detect the case where nothing is actually relevant
This lesson sits in Embeddings and Vector Search, part of Retrieval-Augmented Generation from Scratch. It assumes what came before it and leads directly into the next lesson in the module.
In this lesson you will:
- Compare cosine, dot product, and euclidean distance
- Calibrate a relevance threshold on real queries
- Detect the case where nothing is actually relevant
Resources
Your notes will live here
Note taking is not available yet. Nothing you type would be saved, so the tab stays read-only for now.