| Embedding |
A numeric vector representing the meaning of a text or image. |
| Model |
A SigLIP 2 checkpoint loaded from a local directory (usually the toolkit cache). |
| Checkpoint |
The model weights, configuration, preprocessor and tokenizer files. |
| Embedding space |
The model + preprocessing identity that makes vectors comparable. |
| Manifest |
The embedding_space header stored in an export. |
| Export |
A portable JSONL/pickle file with a header and records. |
| Record |
One vector plus optional id and metadata. |
id |
Stable record identifier; needed for incremental updates and deduplication. |
| Metadata |
Optional JSON data: path, filename, dimensions, EXIF, timestamps, thumbnail path. |
| Thumbnail |
An optional derived image; never embedded in the export. |
| Normalization |
L2 normalisation of embeddings; the toolkit default. |
| Cosine similarity |
Dot product of normalised vectors; the toolkit reports cosine distance in ranking. |
| L2² / euclidean |
Squared/euclidean distance; lower is better. |
| RRF |
Reciprocal Rank Fusion; combines ranked lists without comparing raw scores. |
| Zero-shot classification |
Classify against candidate labels without task-specific training. |
| Template |
Prompt wrapping each label, e.g. a photo of a {}. |
| Fine-tuned classifier |
A model with a trained classification head and its own labels (ImageClassifier). |
| Logits |
Raw model scores before softmax; available via zero_shot_logits. |
| Toolkit home |
Directory containing config.json and the model cache. |
| Model cache |
Downloaded/imported checkpoints under the toolkit home. |
| Revision |
The immutable model revision/hash recorded in the manifest. |
| Preprocess |
The image/text preprocessing contract that must match for vectors to be comparable. |
| Incremental export |
Update only new/modified files and drop deleted ones; needs ids/metadata. |
| Exporter |
The class/CLI that scans images and writes an export. |
| Embedder |
The class that turns texts/images into vectors and runs zero-shot classification. |
| Filters |
Composable metadata predicates (field, matches, to_dict, from_dict). |
| CLI |
The siglip2 command. |
| REST |
The optional FastAPI server (siglip2 serve). |
| Toolkit |
The library, CLI and REST surfaces over the same core. |