| # Generation: gemini (API) or local (Qwen on CPU). auto = Gemini if GEMINI_API_KEY is set. | |
| GENERATION_BACKEND=gemini | |
| GEMINI_API_KEY= | |
| GOOGLE_API_KEY= | |
| GEMINI_MODEL=gemini-2.0-flash | |
| SKIP_LOCAL_LLM_HUB_DOWNLOAD=1 | |
| LOCAL_EMBEDDING_MODEL=all-MiniLM-L6-v2 | |
| LOCAL_LLM_MODEL=Qwen/Qwen2.5-1.5B-Instruct | |
| TASK_A_REVIEWS_EMBEDDED=data/task_a_reviews_embedded.jsonl | |
| TASK_A_RAG_TOP_K=5 | |
| TASK_A_MAX_TOKENS=1024 | |
| TASK_A_TEMPERATURE=0.35 | |
| TASK_B_EMBEDDED_CATALOG=data/business_catalog_embedded.jsonl | |
| TASK_B_LLM_CANDIDATE_CAP=6 | |
| TASK_B_TEMPERATURE=0.2 | |
| TASK_B_MAX_OUTPUT_TOKENS=256 | |
| # llm (Gemini or local Qwen) or retrieval — retrieval is instant (no API) | |
| TASK_B_RANK_MODE=llm | |
| TASK_B_FAST_RANK= | |
| STARTUP_PREWARM=all | |
| SKIP_STARTUP_PREWARM= | |
| HF_TOKEN= | |