env.example

MCP_AUTH_TOKEN=change-me-to-a-long-random-string

# Local default is ./cache. Dockerfile overrides this to /tmp/cache.
# CACHE_DIR=./cache
# MAX_CHARS_PER_PAPER=8000
# EXTRACT_IMAGES=1
# MAX_IMAGES_PER_PAPER=5

# Passage retrieval:
# bm25 = local only; embedding = embeddings only; hybrid = BM25 + embeddings when configured.
# RETRIEVAL_METHOD=hybrid
# RETRIEVAL_TOP_K=4
# DashScope / Qwen embeddings via OpenAI-compatible endpoint.
# If EMBEDDING_API_KEY is omitted, DASHSCOPE_API_KEY or LLM_API_KEY is used.
EMBEDDING_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
# EMBEDDING_API_KEY=sk-xxx
EMBEDDING_MODEL=text-embedding-v4
# text-embedding-v4 supports up to 10 text inputs per synchronous request.
# EMBEDDING_BATCH_SIZE=10

# Optional image OCR / figure understanding.
# OCR_IMAGES=0
# ANALYZE_IMAGES=0
# VISION_IMAGES=0
# MAX_ANALYZED_IMAGES=3
# VISION_MODEL=gpt-4o-mini

# Optional LLM semantic validation for perovskite regex false positives.
LLM_BASE_URL=https://api.deepseek.com
LLM_API_KEY=sk-xxx
LLM_MODEL=deepseek-chat
# PEROVSKITE_LLM_FILTER=1