| # Copy this file to .env and fill in your values. | |
| # .env is git-ignored — never commit secrets. | |
| # ── Hugging Face ──────────────────────────────────────────────────────────── | |
| # Required to download gated models (Qwen3, SDXL-Turbo, Whisper…). | |
| # Get yours at https://huggingface.co/settings/tokens | |
| HF_TOKEN=hf_... | |
| # ── Feature flags ─────────────────────────────────────────────────────────── | |
| # VN_MOCK=1 → fully mocked (no models needed, default for --mode mock) | |
| # VN_MOCK=0 → real backends (install the relevant uv extras first; default for --mode prod/debug) | |
| # | |
| # CLI modes (app.py --mode <mode>): | |
| # --mode mock → force VN_MOCK=1 (no models, fast iteration) | |
| # --mode prod → force VN_MOCK=0 (real AI backends) | |
| # --mode debug → live monitor + end-of-session report; VN_MOCK comes from THIS file | |
| # → set VN_MOCK=0 here to debug real models | |
| # → set VN_MOCK=1 here to debug with mock (UI/flow testing) | |
| VN_MOCK=1 | |
| # LLM runtime backend when VN_MOCK=0. | |
| # llamacpp (default, local) | transformers (HF Space / ZeroGPU) | |
| VN_LLM_BACKEND=llamacpp | |
| # Force the gr.Blocks MVP UI instead of the custom VN frontend. | |
| GRADIO_MVP_UI=0 | |
| # Path to a .jsonl file for Open-Trace logging. Leave empty to disable. | |
| # VN_TRACE=runs/my-session.jsonl | |
| # ── Model registry ────────────────────────────────────────────────────────── | |
| # LLM (transformers backend) | |
| VN_LLM_REPO=Qwen/Qwen3-14B | |
| # LLM (llama.cpp backend) | |
| VN_LLM_GGUF_REPO=Qwen/Qwen3-14B-GGUF | |
| VN_LLM_GGUF_FILE=Qwen3-14B-Q4_K_M.gguf | |
| # Image generation | |
| VN_IMAGE_MODEL=stabilityai/sdxl-turbo | |
| # Optional: HF repo or local path to your anime-style LoRA (leave blank for none) | |
| VN_IMAGE_LORA= | |
| # Whisper model size: tiny | base | small | medium | large-v3 | large-v3-turbo | |
| VN_WHISPER_SIZE=large-v3-turbo | |
| # ── Generation / budget params ─────────────────────────────────────────────── | |
| # Max tokens fed to the LLM per turn ("Thousand Token Wood") | |
| VN_CTX_BUDGET=3500 | |
| # Number of recent turns kept verbatim in context | |
| VN_RECENT_K=6 | |
| # Compact memory every N turns (also triggers when over budget) | |
| VN_COMPACT_EVERY=8 | |
| # Image resolution in pixels (square) | |
| VN_IMAGE_SIZE=768 | |
| # SDXL-Turbo diffusion steps (1–4) | |
| VN_IMAGE_STEPS=2 | |
| # Max NPCs on stage at once | |
| VN_MAX_PRESENT=3 | |
| # ── Paths (defaults are fine; override only if you need a different layout) ── | |
| # VN_MODELS_DIR=./models | |
| # VN_MEMORY_DIR=./memory | |
| # VN_RUNS_DIR=./runs | |
| # VN_CACHE_DIR=./.cache/images | |
| # ── Art style (override to change the global visual aesthetic) ─────────────── | |
| # VN_STYLE_BASE=Japanese anime illustration, soft cel shading, clean linework, ... | |
| # VN_NEGATIVE=text, watermark, signature, lowres, blurry, deformed hands, ... | |