smart-image-search（智能图像搜索）

Search Image

Overview

Handle image-search requests triggered by phrases like search image X or find a picture of X. Default behavior for this skill:

• Search multiple engines, not just one source
• Prefer the most relevant result, not the literal first result
• Download the best match locally, then send it as an attachment when possible
• If confidence is weak, provide several candidates instead of pretending one result is perfect

Engines

Use as many available sources as practical, in this priority order:

• Bing Images
• Baidu Images
• Sogou Images

Notes:

• Bing is usually the most parseable server-side source
• Baidu may return a security-verification page; treat it as a soft failure
• Sogou can provide extra recall when Bing drifts or lacks meme-style results

Workflow

1. Extract the search query after the trigger phrase.
2. Parse lightweight parameters first. Read references/parameters.md.
3. Detect intent on the cleaned core query. Read references/intent-routing.md when tuning or debugging routing.
4. Normalize the cleaned core query, but keep the original wording too.
5. Detect subject subtype when relevant, especially under official: campus, emblem, mascot, poster.
6. Search across multiple engines.
7. Collect several candidate image results instead of only one.
8. Rank candidates by relevance with intent-aware weighting, source trust, and subtype-aware boosts/penalties.
9. Classify confidence. Read references/confidence.md when tuning thresholds.
10. If confidence is high, download and send the best match.
11. If confidence is medium, prefer 2-3 candidates.
12. Apply quality filtering before final send. Read references/quality-filtering.md when tuning thresholds.
13. If confidence is low or downloading fails, return top candidates plus search links.

Parameters

Support lightweight natural-language parameters inside the query. Examples:

• search image taylor swift 3 images
• search image acme logo official
• search image funny cat meme
• search image aurora wallpaper 4k landscape
• search image anime girl avatar 2 images

Supported parameters:

• Count: 1 to 5 images
• Intent override: official, meme, avatar, wallpaper
• Orientation preference: landscape, portrait
• Quality preference: hd, 4k, high resolution

Parse these first, then search the cleaned core entity query.

Intent routing

Classify the request before ranking:

• meme: meme / reaction image / funny image / emoji style requests
• official: official / logo / emblem / mascot / poster / brand identity
• portrait: people, characters, or general image requests
• wallpaper: wallpaper / hd / high resolution / 4k
• avatar: avatar / profile picture / icon

When official is selected, also infer a subtype when possible:

• campus: campus / gate / building / landscape / map
• emblem: emblem / logo / crest / mark / badge
• mascot: mascot / character / brand character / official character / IP character
• poster: poster / promotional art / campaign art

Intent affects ranking:

• meme prefers meme and funny-image pages
• official prefers official domains, institutional pages, and reference pages
• portrait prefers representative images
• wallpaper prefers larger image results
• avatar prefers square or icon-like results

Relevance policy

Do not blindly send the first result. Use references/relevance.md and follow these rules. For official intent, also read references/official-sources.md and references/official-whitelist.md when tuning or debugging official-result ranking. Use references/entity-consistency.md to prevent official-looking but wrong-entity matches. Use references/entity-gating.md when the query contains multiple strong entities and partial matches must be demoted to fallback-only.

• exact query match is best
• token matches in title/page/image URL matter
• semantically related domains help
• obvious news drift, spammy pages, or generic stock images should rank lower
• weak modifiers like image, photo, meme, avatar, and wallpaper should not outweigh the core entity
• in official mode, prefer trusted domains over visually appealing aggregators
• in emblem mode, prefer brand/identity signals and penalize maps, personal homepages, and generic banners
• in mascot mode, prefer official character/IP wording and penalize generic news coverage when cleaner source pages exist

If confidence is weak, prefer sending 2-3 candidates or links rather than a low-quality single answer. Read references/confidence.md for the high / medium / low decision rule.

Output rules

• Be brief and lead with the result.
• If the image was sent successfully, say so in one short sentence.
• If the image cannot be sent directly, do not pretend the search succeeded.
• When falling back, provide clickable search links immediately.
• For meme or mascot queries, optimize for likely intent match, not formal source prestige.

Ready-made search URLs

Use these patterns:

• Bing Images: https://www.bing.com/images/search?q=<urlencoded_query>
• Baidu Images: https://image.baidu.com/search/index?tn=baiduimage&word=<urlencoded_query>
• Sogou Images: https://pic.sogou.com/pics?query=<urlencoded_query>

If needed, use scripts/build_image_search_urls.py to generate encoded URLs safely.

Scripts

Build search URLs

Run:

python3 scripts/build_image_search_urls.py "official mascot"

Parse the most relevant image automatically

Run:

PYTHONPATH=scripts python3 scripts/search_best_image.py "cat meme"

The script prints JSON containing:

• parsed parameters
• cleaned search_query
• per-engine status
• normalized query info
• detected intent
• confidence
• best_image
• top_candidates

Download the best image for direct sending

Run:

PYTHONPATH=scripts python3 scripts/fetch_best_image.py "cat meme"

The script:

• searches Bing, Baidu, and Sogou
• ranks candidates by relevance
• computes confidence
• downloads the best match to /home/mumu/clawd/tmp/search-image/
• prints JSON with path, image_url, engine, score, why, and confidence info

Download 2-3 candidates when confidence is not high

Run:

PYTHONPATH=scripts python3 scripts/fetch_candidate_images.py "cat meme" 3

Use this when confidence is medium or low. The script downloads multiple top candidates and returns their local file paths for sending.

Quality filtering

Use scripts/image_quality.py through the fetch scripts. Default checks now include:

• reject obvious site assets / logos
• reject too-small files
• reject too-small dimensions
• penalize thumbnail-style URLs
• adjust scoring by intent

Legacy first-result scripts

These remain available for debugging or quick comparison:

• scripts/search_first_image.py
• scripts/fetch_first_image.py

Prefer the best-image scripts in normal operation.

Send-image handoff

After fetch_best_image.py succeeds, send the downloaded file as an attachment.

Preferred handoff:

• Use the local file path returned in path
• Use the best-match result, not the literal first result
• If confidence is low but still usable, mention it briefly or send multiple candidates instead
• If attachment sending fails, fall back to sending candidate image URLs and search links

Suggested flow:

1. Run PYTHONPATH=scripts python3 scripts/fetch_best_image.py "<query>"
2. Respect parsed parameters first: count / intent override / orientation / high-res hints
3. If confidence is high and JSON returns ok: true, use the returned local path for image sending
4. If confidence is medium, run fetch_candidate_images.py and send the requested count or 2-3 candidates
5. If confidence is low, send search links plus candidate links or candidate images
6. If JSON returns ok: false, send the search links and top candidate URLs instead

Fallback template

Use this style when direct sending fails:

Result: I could not get a single high-confidence direct image, so here are the best candidates.
Bing Images: <url>
Baidu Images: <url>
Sogou Images: <url>
Candidate 1: <url>
Candidate 2: <url>
Candidate 3: <url>