LAB 3

Embedding Model Impact

Same document, same query, three different embedding models. The model you choose changes what gets retrieved, and therefore what the LLM can answer.

📏Read the order, not the number. Each model has its own scale. Ada-002 packs almost everything into the high numbers, v3 spreads scores lower and wider. A higher percentage here does not mean a better model. Scores are only comparable within one column, never across columns. To judge a model, ask: did it put the right chunk at #1?

Ada 002

Legacy (2022). Baseline quality.

1536D$0.10 / 1M tokens

v3 Small

Latest small. 5× cheaper than Ada, better quality.

1536D$0.02 / 1M tokens

v3 Large

Highest quality. Best for nuanced semantic tasks.

3072D$0.13 / 1M tokens