- The Case for Calibrated Hybrid Retrieval
RRF is a useful default for hybrid search, but it throws away score magnitude. Calibrated retrieval gives us a cleaner path.
9 min - PyLate Late Interaction Retrieval
PyLate makes ColBERT style late interaction practical in Python. Better than bi encoders for precision, costlier to store. Here is when it is worth it.
7 min - Matryoshka Representation Learning
MRL trains embeddings so any prefix of the vector is independently useful. How the multi-scale loss works, which models support it, and how to use it.
10 min - Model2Vec Fast Static Text Embeddings
Model2Vec distills a sentence transformer into a fast static lookup table. How it works, where it fits, and where it falls short.
6 min - Fine-Tuning Embeddings with Contrastive Learning and MRL
Contrastive Learning and Matryoshka Representation Learning.
4 min - Cosine Similarity Limitations in Vector Search
We blindly use cosine similarity for everything in vector search. Here is why that's a bad idea, and what you should do instead.
7 min - Long Context Is Not A Retrieval Strategy
Longer context windows help, but they do not replace retrieval, ranking, and context design.
7 min - LLM Evals Became the New Unit Test
Why evaluation became a core skill for building reliable LLM applications.
9 min