Principle:Cohere ai Cohere python Embedding Type Selection
| Metadata |
|---|
| Cohere Python SDK |
| Cohere Embed Types |
| NLP, Embeddings, Configuration |
| 2026-02-15 14:00 GMT |
Overview
A configuration pattern for selecting embedding input purposes and output formats to optimize vector representations for specific downstream tasks.
Description
Embedding Type Selection involves two orthogonal choices: the input type (how the text should be encoded) and the embedding type (what numerical format the output vectors use). The input type tells the model the intended use -- search_document optimizes for retrieval targets, search_query optimizes for queries, classification and clustering have their own optimizations. The embedding type controls precision and storage tradeoffs: float (full precision), int8/uint8 (quantized for storage efficiency), binary/ubinary (extreme compression for approximate search).
Usage
Always specify input_type when embedding -- using the wrong type degrades retrieval quality. Match search_document for indexing and search_query for queries. For the same search system, always use the same model and consistent input types. Choose embedding_types based on your storage/accuracy tradeoff needs.
Theoretical Basis
Asymmetric embedding models use different encoding strategies for different input roles. This follows the dual-encoder architecture where document and query representations are optimized for cross-similarity rather than self-similarity. Quantized embeddings (int8, binary) approximate the full-precision vectors using fewer bits per dimension, trading accuracy for storage efficiency.