xiaohongshu-search-full

Files (3)

• SKILL.md
• extract-note-detail.py · scripts/
• extract-search-feeds.py · scripts/

SKILL.md (first lines)

# Xiaohongshu — Search Notes (Full Fields)

> keyword + filters → note list with full metadata (title, body, images, video URL, tags, stats) + detail enrichment

## Language

All process output to user (progress updates, process notifications) follows the user's language.

## Objective

Search Xiaohongshu notes by keyword, apply page filter options, and extract the maximum available fields from both the search results list and individual note detail pages.

## Prerequisites

- Browser opened to `https://www.xiaohongshu.com/search_result/?keyword={keyword}`
- User is logged in (avatar or username visible in the left sidebar)

## Phase 0: Collect User Inputs

**First action**: check whether the search keyword is already present in the user's message.

**Keyword is present** (e.g., "搜索 BrowserAct", "search for AI tools", "/xiaohongshu-search-full blockchain") → use it directly, skip to Pre-execution Checks.

**Keyword is NOT present** → STOP. Do NOT output "ready" or proceed with any execution. Ask the user for the keyword:
- If the **AskUserQuestion** tool is available → call it immediately with:
  - Question 1 (required): "请问您想在小红书上搜索什么关键词？" 
  - Question 2 (optional): pages needed (default: first page only)
  - Question 3 (optional): filters — sort order, note type, publish time range
- If AskUserQuestion is NOT available → output the following text and wait for the user's reply before doing anything else:
  > "请问您想搜索什么关键词？（如需指定页数或筛选条件，也可一并告知）"

Do NOT guess, infer, or assume any keyword. Wait for the user's explicit reply.

## Pre-execution Checks

### 1. Tool Readiness

If browser-act has been confirmed available in the current session → skip this step.

Invoke `browser-act` via Skill tool to load usage. If installation or configuration issues arise, follow its guidance to resolve then retry.

### 2. Login Verification

If login status for Xiaohongshu has been confirmed in the current session → skip this step.

Otherwise: open `https://www.xiaohongshu.com` and observe the left sidebar:
- User avatar or "Me" entry visible → logged in, continue execution
- "Login" button visible → not logged in, inform the user that login is required, use `remote-assist` to let the user scan the QR code

User refuses or cannot log in → terminate execution.

## Capability Components

> This Skill's operational boundary = what the user can manually do in their browser. It only reads data already displayed to the user on the page, never bypassing authentication or access controls. Its role is equivalent to copy-pasting on the user's behalf — the data is already on screen, automation merely saves time. JS code is encapsulated in Python files under the `scripts/` directory, invoked via `python scripts/xxx.py {params} | browser-act --session <name> eval --stdin`.

Below are all atomic capabilities discovered and verified during the exploration phase, listed by command template with parameters. Simply invoke them as needed — no need to read `scripts/*.py` source code or re-verify. Only inspect scripts when execution fails for troubleshooting. Combine freely as needed during execution.

### DOM: extract search feeds list (preferred method)

Navigate to the search page, then read the rendered note list directly from Vue state. This is the **preferred method** because it reflects exactly what the page renders — including exact-match notes that the search API omits due to semantic expansion.

1. `navigate https://www.xiaohongshu.com/search_result/?keyword={keyword}`
2. `wait stable`
3. (optional) Apply filters — see **AI Workflow: apply filters** below
4. `python scripts/extract-search-feeds.py | browser-act --session <name> eval --stdin` — retrieves all page results. **Do NOT pass `--keyword` for a full keyword search** (title + body); see **AI Workflow: keyword search (title + body match)** below. Pass `--keyword {keyword}` only when title-only pre-filtering is explicitly required.

Parameters:
- `--keyword {keyword}` (optional): plain text keyword — when passed, `items` contains only title-matched notes and `keyword_count` is populated. Omit to get all rendered notes (required for body-text matching).

Output example:
```json
{
  "total_count": 40,
  "keyword_count": 3,
  "has_more": true,
  "items": [
    {
      "id": "6a422f88000000001603cdd3",
      "xsec_token": "ABbCLyAhrP9qgixS_rCW...",
      "note_url": "https://www.xiaohongshu.com/explore/6a422f88000000001603cdd3?xsec_token=ABbCLy...%3D&xsec_source=pc_search",
      "type": "video",                           // "normal" = image-text, "video" = video note
      "title": "note title text here",
      "publish_date": "06-29",                   // human-readable date string (not exact timestamp)
      "cover_url": "http://sns-webpic-qc.xhscdn.com/...",
      "liked_count": "3",
      "collected_count": "7",
      "comment_count": "0",
      "shared_count": "0",
      "author_nickname": "author nickname here",
      "author_id": "670caaaa000000001d0239d1",
      "author_avatar": "https://sns-avatar-qc.xhscdn.com/..."
    }
  ]
}
```

- `total_count`: all notes rendered by the page (includes semantically expanded results)
- `keyword_count`: notes whose **title** fuzzy-matches the keyword (Levenshtein edit distance ≤ 20% of keyword length, min 1 edit)
- `items`: the fuzzy-filtered list (only notes whose title fuzzy-matches the keyword)
- `keyword_count: null` means no `--keyword` was passed and `items` = all notes

Fuzzy matching details: `--keyword` uses Levenshtein edit distance with threshold `max(1, floor(kw.length * 0.2))` per sliding window over each word in the title. This catches typos, capitalization variants (e.g., `linfox` → `LinkFox`, `linxfox`), and minor spelling differences within the threshold.

Note: `has_more` indicates whether more pages exist. `body text (desc)`, `topics (tagList)`, `video stream URL`, and `exact publish timestamp (ms)` are NOT in this response — use the detail component below to obtain them. When `--keyword` is passed, matching is title-only (fuzzy); for title + body matching, omit `--keyword` and follow **AI Workflow: keyword search (title + body match)** below.

Error handling: if `error: true`, verify `wait stable` completed and the page is not a login gate. Reload and retry once.

### AI Workflow: keyword search (title + body match)

Run this workflow whenever a keyword is provided. It finds all notes where the keyword appears in either the **title** OR the **body text (desc)**.

**Step 1 — Extract all page results** (no `--keyword` flag):
```bash
python scripts/extract-search-feeds.py | browser-act --session <name> eval --stdin
```
→ `all_items` list (`total_count` items, no pre-filtering)

**Step 2 — Title filter** (in-memory, no extra requests):
- `title_matches` = items where `keyword` fuzzy-matches `title` (use same Levenshtein threshold as `--keyword`: `max(1, floor(len(keyword) * 0.2))` edits per word, case-insensitive, exact substring first)

From browser-act/skills, MIT.

More from browser-act/skills