Server-side search over the Hub's pre-built index. Narrow it with the checkboxes —
searching full document text is slower but far more thorough.
Paginate straight through the corpus in dataset order.
Note: the
Note: the
/rows endpoint ignores column
projection and always returns all 19 columns, so each page carries the full
html_content + markdown_content of every row — roughly
1.2 MB per row on this corpus. Page sizes are kept small on purpose.
Search results stay light because /search does honour
columns.
Numeric distributions and null counts for every column, computed by the Viewer.
How this works
This is a Static Space: there is no server and no Python runtime.
Every query is an anonymous HTTP call to the Hugging Face Dataset Viewer
(datasets-server.huggingface.co), which does the filtering, searching and
stats server-side against the indexed parquet. That's why it runs free and stays fast
on a 683 MB corpus that could never fit in this page.
Schema quirks handled for you
- Column names are Vietnamese and several differ only by a slash or a space
(e.g.
Cơ quan ban hành/ Chức danh / Người ký). Every control here is built from the live schema, so nothing is hardcoded. - Dates are plain
DD/MM/YYYYstrings, never date types — so year filtering matches on the trailing 4 characters instead of sorting lexically. - Missing values are the literal string
"...", not null or empty — so "present" checks test for that sentinel. Loại văn bảnhas case near-duplicates (Nghị quyếtvsNghị Quyết) that would otherwise split your result counts. Document type filters are case-insensitive by default to merge them.Phần hết hiệu lựcis effectively empty (null for 61,978 of 61,979 rows) and is hidden.markdown_contentis machine-converted from HTML, so it carries artifacts (\_\_\_\_rules, a trailing/.'page marker) that are stripped before display.
Cross-reference
Every row's id is the ItemID= query param of its url, so each
document resolves to a canonical page on vbpl.vn. The viewer links straight to it.
Not included
html_contentandmarkdown_contenttogether are ~98% of the file; they are fetched only when you open a single document.- The corpus has exactly one Parquet row group, so row-group-level pruning does nothing — the Viewer indexes columns individually instead, which is why this stays responsive.