Video RSS Aggregator (Qwen 3.5 Vision, 4-bit)

This project has been rebuilt around Qwen 3.5 multimodal models and a strict local VRAM budget.

• Inference backend: Ollama (Windows-native, no WSL required)
• Default models: Qwen 3.5 4-bit (q4_K_M) tiers
• Storage: SQLite (.data/vra.db)
• API: FastAPI

Design Goals

• Use Qwen 3.5 vision-capable small models for summarization quality.
• Keep total app VRAM use within 8GB by default.
• Prefer 4-bit model variants for quality/efficiency balance.
• Keep setup simple for Windows users (no WSL).
• Use scene-aware frame extraction with timeline coverage for better visual context.

Requirements

• Python 3.11+
• Windows 10/11
• NVIDIA GPU with at least 8GB VRAM
• Ollama installed on Windows: https://ollama.com/download/windows
• ffmpeg and ffprobe on PATH

Quick Start (Windows)

python -m venv .venv
.venv\Scripts\activate
pip install -e .

Run bootstrap (auto-pulls configured models if missing):

python -m vra bootstrap

Start server:

python -m vra serve --bind 127.0.0.1:8080

Then open http://127.0.0.1:8080/ for the guided installation + configuration GUI. The setup page includes one-click diagnostics for Python, FFmpeg/FFprobe, yt-dlp, and Ollama reachability.

4-bit Model Defaults

Default model priority:

1. qwen3.5:4b-q4_K_M
2. qwen3.5:2b-q4_K_M
3. qwen3.5:0.8b-q8_0 (safety floor when smaller than 2B is needed)

Each processing job selects one model up front based on configured VRAM budget, current runtime VRAM usage, and workload size (transcript + frames). The selected model is pinned for that job; there is no mid-processing model fallback.

Video Processing Intelligence

• Scene-aware frame candidate extraction (ffmpeg scene score) to catch shot changes.
• Uniform timeline sampling fallback/fill to keep temporal coverage when scene cuts are sparse.
• Deduplication by frame content hash before final frame set is sent to the model.
• Model preselection per job using VRAM headroom and estimated per-request overhead.
• SQLite runs in WAL mode with tuned pragmas for better concurrent read/write stability.

Runtime Commands

python -m vra bootstrap
python -m vra status
python -m vra verify --source "https://www.youtube.com/watch?v=dQw4w9WgXcQ"
python -m vra benchmark --source "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

benchmark compares scene_aware vs uniform_only extraction on the same source, reports frame uniqueness metrics, and (by default) runs both through summarization to show latency and output-shape differences.

Environment Variables

Variable	Default	Description
BIND_ADDRESS	127.0.0.1:8080	API bind address
API_KEY	(none)	Optional bearer/API-key auth
VRA_STORAGE_DIR	.data	Download/frame/subtitle storage
VRA_DATABASE_PATH	.data/vra.db	SQLite database path
VRA_OLLAMA_BASE_URL	http://127.0.0.1:11434	Ollama API base URL
VRA_MODEL_PRIMARY	qwen3.5:4b-q4_K_M	First-choice model
VRA_MODEL_FALLBACK	qwen3.5:2b-q4_K_M	Second-priority model
VRA_MODEL_MIN	qwen3.5:0.8b-q8_0	Lowest-priority model
VRA_AUTO_PULL_MODELS	true	Pull missing models automatically
VRA_VRAM_BUDGET_MB	8192	Max VRAM budget across the app
VRA_MODEL_SIZE_BUDGET_RATIO	0.75	Share of budget for base model weight
VRA_MODEL_SELECTION_RESERVE_MB	768	VRAM safety reserve kept free during model selection
VRA_CONTEXT_TOKENS	3072	Context window per request
VRA_MAX_OUTPUT_TOKENS	768	Summary output cap
VRA_MAX_FRAMES	5	Max sampled frames per source
VRA_FRAME_SCENE_DETECTION	true	Enable scene-aware frame selection
VRA_FRAME_SCENE_THRESHOLD	0.28	Scene change sensitivity (ffmpeg scene score threshold)
VRA_FRAME_SCENE_MIN_FRAMES	2	Minimum detected scene frames before blending with uniform sampling
VRA_MAX_TRANSCRIPT_CHARS	16000	Subtitle transcript cap
VRA_RSS_TITLE	Video RSS Aggregator	RSS title
VRA_RSS_LINK	http://127.0.0.1:8080/rss	RSS self-link
VRA_RSS_DESCRIPTION	Video summaries	RSS description

API

• GET / (GUI setup + configuration workspace)
• GET /health
• GET /setup/config
• GET /setup/diagnostics
• POST /setup/bootstrap
• GET /runtime
• POST /ingest
• POST /process
• GET /rss?limit=20

Notes

• GUI setup/configuration workspace is available at / when the server is running.
• This version is optimized for local, Windows-native operation first.