Extract Text Embedding (BGE)

How it fits

Typical backends

Caveats

• Hard constraintThe bound checkpoint must be a BGE-family sentence-transformers snapshot. The embedding dimension is fixed by the bound checkpoint, so binding a different-width one produces vectors of a different width that an existing index will reject.
• Hard constraintThe checkpoint is a bound file artifact, not a download. Bind a BGE sentence-transformers snapshot (config.json + modules.json + weights + tokenizer + 1_Pooling) into the bound slot before deploy; nothing is fetched from the hub at runtime.
• Parameter interactionquery_prefix should be set to the BGE retrieval instruction only for query-side text and left empty for passages; mismatching the two sides degrades retrieval quality.
• I/O contractAn empty String short-circuits to a zero-length Embedding. Feeding that into a similarity stage breaks the math, so filter empty inputs upstream.
• Parameter interactionnormalize produces unit-length vectors so cosine similarity equals the inner product; keep it consistent across the index and the queries.
• Fallbackdevice requesting CUDA resolves to CPU when no compatible accelerator is visible at startup; the smaller BGE variants run acceptably on CPU.
• State lifetimeThe bound checkpoint, device, prefix, and normalisation are read once at startup; changing them requires a redeploy.