The Quirks of Embeddings

Draft — not listed anywhere. Reachable only at this URL.

Outline:

• Asymmetric embeddings — why input_type exists, and what breaks when you index documents and queries the same way.
• Chunking — fixed-size vs. semantic vs. structural, and why chunk boundaries silently decide your recall.
• Dimensionality — Matryoshka embeddings, truncation, and the cost/accuracy curve.
• Normalization — when cosine and dot product are the same operation and when they very much are not.
• Distance is not confidence — 0.82 is not "82% relevant", and thresholds don't transfer between models.
• Domain drift — the general-purpose model that never learned your product names, and what to do about it.
• Multilingual and code — where the general models fall apart.
• Evaluating — building a retrieval eval set before you pick a model, not after.