duckdb-xpath

DuckDB extension for XPath queries on XML columns. Uses simdxml for parsing.

Functions

Function	Return	Description
xpath_text(xml, xpath)	VARCHAR	Text of first matching node
xpath_list(xml, xpath)	VARCHAR	JSON array of all matching text
xpath_count(xml, xpath)	BIGINT	Count of matching nodes
xpath_exists(xml, xpath)	BOOLEAN	Whether any node matches
xpath_eval(xml, xpath)	VARCHAR	Scalar XPath expressions (count(), string(), etc.)
read_xml(file, xpath)	Table	Each matching text as a row
xpath_create_index(file)	Table	Build .sxi sidecar index
xpath_drop_index(file)	Table	Remove .sxi sidecar index

SELECT xpath_text(xml, '//invention-title') FROM patents;
SELECT ucid, xpath_count(xml, '//claim-text') AS n_claims FROM patents;
SELECT * FROM read_xml('corpus.xml', '//patent/title');

Benchmark

10K real patent XML documents (1.1 GB, 111 KB avg), release build, compared to webbed (libxml2):

Query	duckdb-xpath	webbed	Ratio
//invention-title text (100% match)	0.005s	3.42s	685x
xpath_exists //claim-text (93% match)	0.025s	3.71s	148x
xpath_count //claim-text (461K total)	0.035s	3.64s	104x
xpath_text //claim-text (93% match)	0.655s	3.75s	5.7x

Simple //tagname queries use a grep-mode fast path (SIMD string search, no XML parsing). Complex queries fall through to structural parsing with buffer reuse across documents.

Corpus files are in test/benchmark/ (git-lfs).

Building

git clone --recurse-submodules <repo>
cd duckdb-xpath
make configure
make release
make test_release

Extension output: build/release/extension/duckdb_xpath/duckdb_xpath.duckdb_extension

LOAD '/path/to/duckdb_xpath.duckdb_extension';

XPath coverage

XPath 1.0: 327/327 libxml2 conformance, 1015/1023 pugixml conformance. XPath 2.0/3.1 in progress upstream in simdxml.

License

MIT OR Apache-2.0