--- title: E5 Embedding Api emoji: 🚀 colorFrom: indigo colorTo: yellow sdk: gradio sdk_version: 6.30.0 python_version: '3.12' app_file: app.py pinned: false license: mit --- Pinned multilingual E5 inference. The existing `/encode` endpoint keeps its list output, prefix handling, empty-input behavior and unit-normalized vectors. The original SentenceTransformer model includes a Normalize module; its encode flag alone does not describe the final output normalization. Model/tokenizer weights are now loaded from the full commit `3d7cfbdacd47fdda877c5cd8a79fbcc4f2a574f3`. `/metadata` does not allocate a GPU. `/encode_pinned(text, kind)` accepts `query` or `passage`, applies that kind's prefix to the exact original text, refuses overlength inputs instead of truncating them and returns normalized FP32 vectors with contract and input hashes. Both inference endpoints share a serial GPU queue. The dependency file selects reviewed Linux x86_64 / CPython 3.12 wheels from this directory's separate uv.lock. Each URL carries its reviewed SHA-256 fragment. Global pip hash mode is avoided because the Spaces installer adds its own requirements; those additions and Gradio OAuth/MCP extras are included in the lock. A requirements hash does not attest the complete host image. The platform's datasets package is included in the solved graph; startup requires pip check to pass before loading model weights. Runtime metadata reports that check, exact package versions and the reviewed source/export hashes. Compare real passage canaries before accepting a deployment. Metadata is an operational check, not cryptographic remote attestation. See https://huggingface.co/docs/hub/spaces-zerogpu and https://huggingface.co/intfloat/multilingual-e5-large for the implementation patterns.