Embeddings cookbook

12 recipes · 78 steps · about 68 minutes of reading

Verify your first embedding works
Copied a tutorial's embedding call. No idea if it's any good.
BeginnerRead 5 minDoing: 20 to 30 minutes6 steps
Choose cosine similarity or dot product for your vectors
Ranking the same two vectors by dot product gave a different top match than ranking them by cosine similarity, and I had no way to know which one to trust.
BeginnerRead 4 minDoing: under an hour6 steps
Add semantic search to your own notes
I know I wrote something down somewhere, but I can't remember the exact words, and keyword search finds nothing.
Beginner to intermediateRead 6 minDoing: about an hour6 steps
Why did swapping your embedding model wreck search quality?
No crash. No error. Search just quietly got worse right after we swapped the embedding model.
Beginner to intermediateRead 5 minDoing: an afternoon6 steps
Fix the cold start problem for a brand-new product's recommendations
I just launched a product with zero purchases, and the 'similar items' box has nothing to show for it.
Beginner to intermediateRead 5 minDoing: an afternoon7 steps
Embed a document longer than your input limit
My embedding call worked fine in every tutorial example, then started failing the moment I pointed it at one of our longest real documents, and I had no idea why.
IntermediateRead 6 minDoing: an afternoon7 steps
Support tickets sorted into topics without hand labeling every one
Thousands of tickets. Nobody has time to read them.
IntermediateRead 6 minDoing: an afternoon8 steps
Untangle five different phrasings of the same question in your dataset
Our support logs have the same question typed five different ways, and exact-match dedup misses every one of them.
IntermediateRead 5 minDoing: an hour5 steps
Pick an embedding model from the MTEB leaderboard, then verify it on your data
The MTEB leaderboard says one model wins, industry blogs repeat the same recommendation, and none of that tells me whether it performs well on my own private documents.
IntermediateRead 6 minDoing: an afternoon7 steps
Find a photo by typing what's in it, using CLIP embeddings
I know I have the photo somewhere, but none of my files or folders are named anything close to what's in it.
IntermediateRead 5 minDoing: about an hour6 steps
Speed up nearest neighbor search with a FAISS index
Instant with a hundred vectors. Unusable with real data.
Intermediate to advancedRead 7 minDoing: an afternoon7 steps
Shrink your vector index's memory footprint with quantization
My local index worked fine in testing and then needed far more memory than I budgeted for once it held real data.
Intermediate to advancedRead 8 minDoing: an afternoon7 steps

Cookbook badge

Finish all 12 recipes to earn this.

Embeddings cookbook

Verify your first embedding works

Choose cosine similarity or dot product for your vectors

Add semantic search to your own notes

Why did swapping your embedding model wreck search quality?

Fix the cold start problem for a brand-new product's recommendations

Embed a document longer than your input limit

Support tickets sorted into topics without hand labeling every one

Untangle five different phrasings of the same question in your dataset

Pick an embedding model from the MTEB leaderboard, then verify it on your data

Find a photo by typing what's in it, using CLIP embeddings

Speed up nearest neighbor search with a FAISS index

Shrink your vector index's memory footprint with quantization

In these roadmaps

Embeddings cookbook

Verify your first embedding works

Choose cosine similarity or dot product for your vectors

Add semantic search to your own notes

Why did swapping your embedding model wreck search quality?

Fix the cold start problem for a brand-new product's recommendations

Embed a document longer than your input limit

Support tickets sorted into topics without hand labeling every one

Untangle five different phrasings of the same question in your dataset

Pick an embedding model from the MTEB leaderboard, then verify it on your data

Find a photo by typing what's in it, using CLIP embeddings

Speed up nearest neighbor search with a FAISS index

Shrink your vector index's memory footprint with quantization

In these roadmaps