Video material finder

Video Material Finder

Find videos by evidence, not by guesswork. Default to Chinese interaction; switch to English when the user writes in English or requests it.

Choose the workflow

• Use Specific video when the user wants a known work, upload, music video, interview, scene, original source, or higher-resolution copy.
• Use Reusable footage when the user wants stock footage, B-roll, public-domain media, Creative Commons media, or material safe to download and edit.
• Run both workflows when the request is ambiguous, then label the two result groups separately.

Before searching, extract the strongest clues: exact title, people, lyrics or dialogue, creator, platform, approximate year, duration, visual landmarks, language, and desired resolution. Generate a compact original-language query and, when useful, one English or Chinese alternate query.

Specific video workflow

1. Search authoritative or original sources first: creator/label/studio accounts, official archives, publisher pages, then reputable platform uploads.
2. Run scripts/search_specific_video.py for YouTube, Bilibili, and optional Brave Video Search. If platform-specific search tools already exist in the environment, use them as additional evidence.
3. If the user supplied a local clip they own or are authorized to analyze, inspect metadata and extract a few keyframes with FFmpeg. Do not upload private media to third-party reverse-search services without permission.
4. Compare title, creator, duration, upload date, frame content, audio clues, and resolution. Do not call a candidate a match from title similarity alone.
5. Return the best match, 2-4 alternatives, why each matches, maximum verified resolution, and a source page link.
6. A viewable video is not automatically downloadable. Offer a direct file only for user-owned, public-domain, or explicitly licensed material. Never bypass DRM, login, geographic restrictions, expiring authorization, or paid access.

Read references/specific-video.md when the evidence is incomplete, multiple copies disagree, or the user asks for downloading.

Reusable footage workflow

1. Search with scripts/search_stock_video.py. Start with Wikimedia Commons; add NASA and Library of Congress for archival material. Enable Pexels/Pixabay when their API keys exist.
2. Prefer exact licensing metadata over site reputation. Keep the provider page together with every file URL.
3. Apply the status gate:
   • reusable: explicit compatible license or public-domain dedication; direct download may be offered.
   • review: rights vary, third-party credits may apply, or the catalog only provides a rights statement; show the link but require human review before publication.
   • blocked: all-rights-reserved, missing source, or incompatible/unknown license; do not present as downloadable stock.
4. Check format, dimensions, orientation, duration, file size, attribution, model/property/trademark restrictions, and commercial-use constraints.
5. Return a short ranked list. Include source_page, direct_download, resolution, license name/link, attribution, usage status, and restrictions.

Read references/providers.md for setup and API behavior, references/licensing.md before making reuse claims, and references/output-format.md for the response schema.

Commands

Probe the local environment without exposing secrets:

python scripts/probe_sources.py

Search reusable footage with zero-key providers:

python scripts/search_stock_video.py "ocean waves" --providers commons,nasa,loc --limit 5 --format markdown

Add commercial stock APIs when configured:

python scripts/search_stock_video.py "城市夜景" --alt-query "city night aerial" --providers all --min-width 1920 --format markdown

Search for a specific video without downloading it:

python scripts/search_specific_video.py "Frank Ocean Endless visual album" --providers youtube,bilibili,brave --limit 5 --format markdown

API keys are optional environment variables: BRAVE_SEARCH_API_KEY, PEXELS_API_KEY, and PIXABAY_API_KEY. Never print or commit their values.

Non-negotiable rules

• Distinguish free to view, free to download, royalty-free, Creative Commons, and public domain. They are not synonyms.
• Treat every direct file URL as coupled to its source page and license evidence.
• Do not scrape sites whose terms prohibit it; use documented APIs or manual links.
• Do not mass-download, mirror, or redistribute stock libraries.
• Do not claim legal certainty. State what the provider says and flag jurisdiction, people, logos, music, artworks, and property as separate rights.
• Ask before downloading a large file when size or destination is unclear.
• Prefer stable provider pages over temporary CDN links in long-lived reports.