catlog22 abdc66cee7 feat: add model download manager with HF mirror support and fix defaults ...

- Add lightweight model_manager.py: cache detection (with fastembed name
  remapping), HF mirror download via huggingface_hub, auto model.onnx
  fallback from quantized variants
- Config defaults: embed_model -> bge-small-en-v1.5 (384d), reranker ->
  Xenova/ms-marco-MiniLM-L-6-v2 (fastembed 0.7.4 compatible)
- Add model_cache_dir and hf_mirror config options
- embed/local.py and rerank/local.py use model_manager for cache-aware loading
- Fix FastEmbedReranker to handle both float list and RerankResult formats
- E2E test uses real FastEmbedReranker instead of mock KeywordReranker

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-03-17 10:29:52 +08:00

6.9 KiB

Raw Blame History

View Raw