# CPU-only OpenAI-compatible embedding endpoint for 1C code search.
# The model cache is persistent on docker-gpu; no credentials are required.

EMBEDDING_CONTAINER_NAME=llm-qwen3-embedding
EMBEDDING_IMAGE=ghcr.io/ggml-org/llama.cpp@sha256:3e8914c1aab600a330ada97c10fb5fb02ff1c15ac39f2ece218db125cf54594e
EMBEDDING_HOST_PORT=8082
EMBEDDING_HF_REPO=Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0
EMBEDDING_SERVED_MODEL_NAME=qwen3-embedding-0.6b
EMBEDDING_CTX_SIZE=8192
EMBEDDING_BATCH_SIZE=1024
EMBEDDING_UBATCH_SIZE=1024
EMBEDDING_THREADS=12
EMBEDDING_PARALLEL=1
EMBEDDING_CPU_LIMIT=8.0
EMBEDDING_MEMORY_LIMIT=4G

# Windows path on docker-gpu.cin.su, mounted into the Linux container.
HOST_LLAMA_CACHE_DIR=Z:/LLM/models/cache/llama.cpp