Revert to stable config: BATCH_TOKENS=512, CONCURRENT=8

Reduce MAX_BATCH_TOKENS for CPU stability

MAX_BATCH_TOKENS=32768, MAX_INPUT_LENGTH=8192

Increase MAX_BATCH_TOKENS to 32768, MAX_INPUT_LENGTH=8192

Increase concurrency limits

Increase concurrency: MAX_CONCURRENT_REQUESTS=8, MAX_BATCH_REQUESTS=16, MAX_BATCH_TOKENS=1024

Upload Dockerfile with huggingface_hub

Reduce memory: max_concurrent=1, max_batch_tokens=512

Try float16 to reduce memory usage

Fix: use TEI 1.9 (1.6 doesn't support Qwen3)

Deploy Qwen3-Embedding-0.6B with TEI

Restore stable PyTorch CPU + 4B (backup while INT8 ONNX upload in progress)

TEI INT8 ONNX + --pooling cls --dtype float32

TEI + INT8 ONNX Qwen3-Embedding-4B (4GB, fits cpu-basic)

Stable: PyTorch CPU-only + Qwen3-Embedding-4B (proven working)

Remove build-time model download; use ORT backend at runtime

Add ONNX Runtime backend option + ST fallback for Qwen3-Embedding-4B

TEI Candle CPU + Qwen3-Embedding-4B fp16

Use CPU-only PyTorch for smaller image & faster inference

Switch to sentence-transformers + Qwen3-Embedding-4B (PyTorch CPU)

Use ONNX ST Qwen3-Embedding-4B model

Try TEI with Qwen3-Embedding-4B

Switch to TEI cpu-1.9 for Qwen3-Embedding-8B

Add Qwen3-Embedding-8B API

initial commit