Search via Google CSE API. Args: date_restrict: Google CSE dateRestrict value, e.g. "y2" (past 2 years), "m6" (pa

Compute similarity between two Chinese titles (0.0 - 1.0).

Write markdown file. Returns True if content changed.

Quote a string if it contains YAML-special characters.

Extract the guidance document title from markdown content. Handles NMPA guidance document patterns: - "附件N" prefix lines (skipped) - Titl

Detect if extracted text is a catalog/list of guidance titles.

Check if the document content plausibly matches the expected title. Also detects the NMPA "attachment mismatch" pattern where the title appea

Detect if extracted text is a draft for comments (征求意见稿). Only official/final versions should be included in the knowledge base. Draft docume

Parse a Markdown table row into a list of cell strings.

Build HTML full text URL from CELEX number.

Fetch page content. Returns (status_code, html_text).

Get the English translation output path for a Chinese fulltext file.

Return True if this is a Markdown table separator row (---|---|...).

Parse YAML front matter from a Markdown file.

Run a git command and return (returncode, stdout, stderr).

Get direct <tr> children of a table (or its <tbody>/<thead>). Unlike ``table.find_all("tr")``, this does NOT recurse into nested tables, whic

Escape angle brackets that VitePress/Vue would interpret as components. Preserves: standard HTML tags, Markdown auto-links <https://...>, HTM

Save validated fulltext to the entry's slug file. Returns True if index was modified.

Extract meaningful keywords from a Chinese title.

Classify a standard into a category based on number and title.

Clean HTML entities from title.

Derive repo output path and regulation from WordPress link URL. Returns (output_path, regulation).

Remove year/revision markers and normalize suffix.

Append a single entry line to the index page. link_prefix: override the relative path prefix for fulltext links. Default "" mean

Run DatabaseComparator on each new_item; keep only new/update items.

Extract text from a table cell, recursively processing nested layout tables.

Use curl-cffi with Chrome TLS impersonation, bypassing proxy for FDA. Akamai bot detection blocks requests from proxy/VPN IPs.

Find LibreOffice binary. Caches result.

Recursively convert a BeautifulSoup element to markdown lines.

Format a table row for a document entry.

Extract the core product/test name by stripping generic suffixes.

Filter search results, return (new_items, domain_filtered, title_filtered).

Check if any of the tag's CSS classes contain cls_fragment.

Return True if line is boilerplate that should not be used as a title.

Detect EUR-Lex layout tables used for numbered/bulleted definitions. Uses only *direct* rows (not recursive) to avoid being confused by neste

Convert a layout table to indented text lines. Only processes *direct* rows to avoid duplication from nested tables. Nested tables inside cel

Load standards from cache file.

Load fulltext markdown for an entry, if available. Checks both {name}.md and {name}.zh.md (NMPA Chinese sources often use the .zh.md extensio

Load translated fulltext for a specific language. Looks for {slug}.{lang}.md or {id}.{lang}.md in the fulltext dir. Returns empty string if n

Heuristic: true titles don't end with descriptive punctuation and are short.

Normalize a Chinese title for comparison (remove spaces, punct variants).

Convert a real data table to Markdown table syntax. Only processes *direct* rows to avoid pulling in nested table content.

Remove L2 codes that only appear as substrings of L3 codes in the text. Problem 1 (tail): text has '02-06-01', broad match picks up '06-01' (vali

Classify a detection against the database. Returns: ('new', description) | ('update', diff) | ('known', '') | ('irrelevant', reason)

Convert .doc (OLE2) to .docx via LibreOffice headless.

Convert a PDF (URL or local path) to Markdown using docling. Args: source: URL or local file path to the PDF doc_id: Documen

Detect insights subcategory from post categories.

Fetch EUR-Lex HTML and parse to markdown. Retries on HTTP 202 (Accepted) which EUR-Lex returns when generating large documents asynchronously

Fetch HTML from URL with caching.

Generate the main news index page.

Generate an English standards sub-page from JSON data.

Generate a Chinese standards sub-page from JSON data.

Fetch all published docs with pagination.

Fetch all published posts with pagination.

Get category slug by ID (cached).

Get entries from JSON data, handling both 'entries' and 'standards' keys.

Get the fulltext markdown file path for an entry.

Get title string, handling both string and dict formats.

Group standards by (l1, l2) categories.

Group NMPA guidance entries by device category using Chinese title keywords.

Check if line is a table separator like `---|---|---`.

Check if line looks like a table data row (contains | but is not a separator).

Normalize a table row to have leading/trailing pipes.

Normalize separator row.

Normalize a Chinese title for fuzzy matching.

Process a single Markdown file: find external images, download them, update references. Returns number of images processed.