vidlizer/env.sample at main · arizawan/vidlizer · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
# Copy to .env and fill in your values.

# ── Provider ────────────────────────────────────────────────────────────────
# openrouter  — cloud inference via OpenRouter (requires API key)
# ollama      — local inference via Ollama (no API key, no cost)
# openai      — any OpenAI-compatible server: LM Studio, vLLM, LocalAI, real OpenAI
PROVIDER=ollama

# ── OpenRouter (cloud) ──────────────────────────────────────────────────────
# API key — get one at https://openrouter.ai/keys
OPENROUTER_API_KEY=sk-or-v1-REPLACE_ME

# Default cloud vision model.
# Cheap + strong:  google/gemini-2.5-flash (~$0.001/run)  |  google/gemini-2.5-flash-lite
# Free tier:       nvidia/nemotron-nano-12b-v2-vl:free  |  google/gemma-3-27b-it:free
# Expensive:       google/gemini-2.5-pro  |  openai/gpt-4o  (warns before running)
OPENROUTER_MODEL=google/gemini-2.5-flash

# ── Ollama (local) ──────────────────────────────────────────────────────────
# Ollama server URL (default: http://localhost:11434)
OLLAMA_HOST=http://localhost:11434

# Default local model — requires 8 GB+ RAM minimum:
#   qwen2.5vl:3b   ~3.2 GB disk / ~5 GB RAM  ★ recommended — 125K ctx, strong JSON, multi-image
#   qwen2.5vl:7b   ~6.0 GB disk / ~9 GB RAM    best quality, needs 10+ GB RAM
#   minicpm-v:8b   ~5.5 GB disk / ~8 GB RAM    strong OCR + reasoning, needs 10+ GB RAM
# Install: ollama pull qwen2.5vl:3b
OLLAMA_MODEL=qwen2.5vl:3b

# ── OpenAI-compatible (LM Studio, oMLX, vLLM, LocalAI, real OpenAI) ─────────
# Base URL for the OpenAI-compatible server (no trailing slash)
#   LM Studio default: http://localhost:1234/v1
#   oMLX default:      http://localhost:8000/v1
#   vLLM default:      http://localhost:8000/v1
OPENAI_BASE_URL=http://localhost:1234/v1
# API key — use "lm-studio" for LM Studio, "not-needed" for vLLM/oMLX, real key for OpenAI
OPENAI_API_KEY=lm-studio
# Model name exactly as shown in the server (required)
# LM Studio example: qwen/qwen2.5-vl-7b-instruct
# oMLX example:      mlx-community/Qwen2.5-VL-7B-Instruct-8bit
OPENAI_MODEL=

# Frame extraction knobs
SCENE_THRESHOLD=0.1       # 0..1, lower = more frames (hybrid: scene-change OR interval)
MIN_INTERVAL=2            # seconds between forced frames (fallback; scene-change fires first)
MAX_FRAMES=60             # HARD CAP — prevents runaway API cost
FRAME_WIDTH=512           # px; higher = more detail, more tokens

# ── Fallback (cross-provider or same-provider) ───────────────────────────────
# FALLBACK_PROVIDER: if set to a different provider, switches provider on failure
#   Same provider: leave blank (auto-detects installed models)
#   Cross-provider: set to  ollama | openai | openrouter
FALLBACK_PROVIDER=
# FALLBACK_MODEL: model to try when primary fails (required for cross-provider)
#   Same-provider example:  qwen2.5vl:3b  |  qwen/qwen2.5-vl-3b-instruct
#   Cross-provider example: qwen2.5vl:7b  (if FALLBACK_PROVIDER=ollama)
FALLBACK_MODEL=
# FALLBACK_BASE_URL: base URL for fallback provider (openai-compat or ollama)
#   Defaults to OPENAI_BASE_URL or OLLAMA_HOST if not set
FALLBACK_BASE_URL=
# FALLBACK_API_KEY: API key for fallback provider
#   Defaults to OPENAI_API_KEY or OPENROUTER_API_KEY if not set
FALLBACK_API_KEY=

# Safety / cost caps
MAX_COST_USD=1.00         # abort mid-run if estimated spend exceeds this
BATCH_SIZE=0              # 0 = auto (one request if possible)
REQUEST_TIMEOUT=600       # per-API-call seconds