Glossary

A classification metric that balances precision and recall into a single number by taking their harmonic mean. Useful when both false positives and false negatives carry cost, and when class distribution is uneven. Ranges from 0 to 1, with 1 being perfect precision and recall.

Performing a task from just a handful of examples rather than a large training dataset. Some embedding models adapt to new tasks with only a few labeled examples, which matters for specialized domains where labeled data is scarce.

Taking a pre-trained model and continuing to train it on a more specific task or dataset. For embedding models, fine-tuning typically means training the model to place similar content close together and dissimilar content far apart. Investing in data quality, task specialization, and fine-tuning for embedding quality counts for much more than the choice of base architecture.

The initial stage of a multi-stage retrieval pipeline, responsible for quickly narrowing millions of candidates down to a smaller set, typically hundreds, for further reranking. First-stage retrieval prioritizes speed and recall over precision, accepting some irrelevant results in exchange for broad coverage. Both dense methods, such as bi-encoder embedding models, and sparse methods, such as BM25, are commonly used at this stage.

A measure with two related meanings in machine learning: hardware FLOPS refers to how many floating point operations a processor can perform per second, measured in GFLOPS, TFLOPS or PFLOPS; model FLOPS refers to the number of floating point operations required for a single forward pass, used to quantify computational cost independent of hardware. Larger models require more FLOPS per inference, and quantization reduces FLOPS by replacing high precision operations with lower precision equivalents.