crispembed 0.16.1

Lightweight embedding + math OCR inference via ggml. Dense, sparse, ColBERT, cross-encoder reranking, and DeiT+TrOCR on-device math equation recognition.

example/README.md

crispembed examples

crispembed runs lightweight text-embedding and OCR/OMR inference on-device via ggml. See feature_parity.dart in this directory for an end-to-end parity harness; the essentials:

Text embeddings

import 'package:crispembed/crispembed.dart';

void main() {
  // Needs the native libcrispembed at runtime (see the package README).
  final model = CrispEmbed('/path/to/model.gguf');
  try {
    final embedding = model.encode('A short document');
    print('dim = ${embedding.length}');
  } finally {
    model.dispose();
  }
}

What else is here

crispembed also exposes sparse (BGE-M3 / SPLADE) and ColBERT multi-vector embeddings, cross-encoder reranking, and the OCR/OMR engines (math_ocr, omr, hmer_ocr, bttr_ocr, text_detect, layout, scan_cleanup) — all imported from package:crispembed/crispembed.dart.