Distill — Turn Any Video into Knowledge

🏆 Built for FutureAI Global Hackathon 2026

Paste a video link, get structured knowledge. Distill transcribes Bilibili videos and uses multiple LLMs in parallel to generate summaries, structured study documents, and mind maps — then lets an independent judge model pick the best output.

🌐 Live at wd-ai.cloud

China's 200M+ exam-prep students (考研/考公) routinely face 3-hour lecture videos on Bilibili and Douyin. Tools like NotebookLM serve YouTube and English content — this audience has nothing equivalent. Watching at 2× speed still leaves zero reviewable notes.

The following were built during the FutureAI Global Hackathon 2026:

• End-to-end Bilibili ingestion — paste a video URL, /api/transcribe downloads the audio via yt-dlp and transcribes it through SiliconFlow's cloud ASR, with duration validation + automatic retry to catch a CDN edge case where downloads silently truncate (see docs/EXPERIENCE.md)
• LLM-as-judge — DeepSeek-v4-pro and Kimi-k2.6 generate in parallel on the same transcript; Doubao-seed-2.0-pro evaluates both on coverage, structure, and factuality and picks a winner, with scores shown to the user
• Next.js 15 frontend — App Router rewrite with TypeScript, supporting both direct-text input and Bilibili-URL ingestion in one interface

All three models are accessed through a single unified gateway (Volcano Ark coding plan), differentiated only by model parameter.

Runs on a dedicated server via docker compose up -d. Simple by design — the focus this hackathon was the product, not the ops.

Browser
  │  HTTP (synchronous request, no SSE yet)
  ▼
Next.js 15 (App Router)
  │
  ├── /api/transcribe → yt-dlp → SiliconFlow ASR → transcript
  │
  └── /api/generate   → LLM Router (embedded in apps/web/lib/)
                              │
                    ┌─────────┴─────────┐
              DeepSeek-v4-pro      Kimi-k2.6
                    └─────────┬─────────┘
                      Doubao-seed-2.0-pro (Judge)
                              │
                         Winner + scores

Full diagram: docs/architecture.md

# 1. Clone
git clone https://github.com/misky530/distill.git
cd distill

# 2. Configure environment
cp .env.example .env.local
# fill in DEEPSEEK/KIMI/DOUBAO keys (Volcano Ark), SiliconFlow key,
# BILIBILI_COOKIES_PATH, YT_DLP_PATH, FFMPEG_PATH

# 3. Install & run
pnpm install
pnpm --filter web dev

No database or queue service is required to run the current feature set — Postgres/Redis/BullMQ are part of the roadmap, not yet wired into the request path.

• Bilibili only — Douyin URL validation exists in code but the download/transcription path is untested.
• Processing is synchronous HTTP; long videos (30min+) may hit request timeouts. Async queueing is on the roadmap, not yet implemented.
• Requires a manually exported Bilibili cookie file (no auto-refresh); see docs/EXPERIENCE.md for why --cookies-from-browser doesn't work on Windows.
• Pro mode (dual-LLM generation + judge scoring) can take 3-5 minutes end-to-end, with only a static status message — no progress bar. This is expected given three sequential/parallel LLM calls, but worth knowing before testing it live.

• Async task pipeline (BullMQ + Redis) with SSE progress streaming, to support 1h+ videos
• Two-tier LLM caching (Redis exact + pgvector semantic)
• Douyin ingestion support
• Observability: Loki + Grafana + distributed tracing
• WeChat Mini Program login for Chinese market
• Migrate LLM orchestration into n8n for visual workflow editing
• Pluggable judge model (swap Doubao → Claude via one config change)

MIT — see LICENSE