Content Review Agent

LangChain-based content moderation orchestration: a deterministic pipeline (wordlists, fingerprints, OCR, etc.) runs alongside an Agent + tool packs; multi-vendor LLM adapters are unified, with optional queues, SQLite audit trails, and metrics. Suited for content safety and UGC governance.

Docs: Layering (L0) · Agent tool packs

Requirements

• Python ≥ 3.9 (3.10+ recommended)
• Local models: Ollama or a compatible inference server

Quick start

python3 -m pip install -e .
cp .env.example .env
# Edit .env and config.yaml as needed (keep secrets in the environment or local .env, not in committed config.yaml)

content-review check "text to review"
content-review check /path/to/image.jpg
content-review check /path/to/video.mp4
content-review file ./a.txt ./b.png
content-review interactive
content-review server --host 0.0.0.0 --port 18080

check infers text / image / video automatically; you do not need --type. The server root serves a simple review page.

If the console script is not on your PATH: python cli.py … (same subcommands).

Optional installs

• pip install -e ".[image]" / ".[video]" — image / video detection extras (see pyproject.toml); the terminal UI (content-review tui) is included in the default dependencies.

Video and FFmpeg

Local video detection needs ffprobe and ffmpeg on PATH (e.g. on macOS: brew install ffmpeg). If they are missing, the pipeline degrades and records labels such as ffprobe_not_found.

Optional Whisper: REVIEW_AGENT_ASR_MODEL (default tiny), REVIEW_AGENT_ASR_LANGUAGE (use zh for Chinese).

Wordlists and knowledge base

• Wordlists: config/wordlists/
• Vector store: put documents under config/knowledge/, then run content-review knowledge ingest-config (or ingest <dir>)

Offline bundle

./scripts/package_bundle.sh   # or: python3 scripts/package_bundle.py

Artifacts land in dist/; follow INSTALL.txt inside the bundle to install.