Vancouver legaltech startup LexSelect says its document-parsing technology outperformed several established platforms in a new benchmark designed around the complexities of legal documents.
The company tested its parsing engine against six competing platforms, including AWS Textract, Reducto, LlamaParse, Docling and Unstructured. The evaluation examined whether each system could find content, correctly identify its role and isolate it precisely enough to support reliable citations.
According to results published by LexSelect, its platform recorded a score of 0.735 for finding and correctly labelling document content, compared with 0.569 for the next-highest competitor, Reducto.
On a stricter “citable” standard—which required systems to capture nearly all of a passage without including neighbouring content—LexSelect scored 0.563. AWS Textract placed second at 0.311.
The benchmark used 24 public-record documents spanning 917 pages and containing 21,660 human-labelled regions. The materials included depositions, trial transcripts, pleadings, agreements, bylaws, scanned insurance forms, a patent and a procurement request for proposals.
Each parser received the same PDFs and was assessed using the same process. LexSelect said the reference regions were drawn, classified and peer-reviewed by human evaluators.
The company led nine of the 12 document categories tested, with particularly large margins on procurement documents, depositions and trial transcripts. It also performed strongly at identifying content such as footnotes rather than simply extracting their words as undifferentiated body text.
That distinction matters in legal AI because the position and function of text can be as important as the text itself. A footnote may qualify a preceding statement, while a clause nested within a contract section may only make sense when read alongside its parent provision.
Precise boundaries are also important when an AI product highlights the passage supporting an answer. Among the regions successfully matched under the standard test, 77 per cent of LexSelect’s selections met the tighter citation requirements, compared with 56 per cent for AWS Textract.
LexSelect cautioned that the benchmark measures content classification and source location—not character-level transcription, the correctness of AI-generated answers or whether a cited passage actually supports a conclusion. Those questions must still be evaluated within the downstream legal workflow.
The company also reported processing speeds of approximately 0.3 seconds per page, compared with 0.4 seconds for Reducto and 2.2 seconds for AWS Textract.
Founded in 2023 by CEO Morgan Maguire, Clio founder and CEO Jack Newton, and intellectual property litigator Scott Foster, LexSelect is building a structured data layer for legal AI. Its technology is designed to preserve exhibits, clauses, subclauses, tables and other relationships across entire documents rather than reducing a filing to a flat stream of text.
LexSelect says that structure can help developers retrieve only the material needed for a task, maintain links to the original source and reduce the number of tokens sent to a large language model.
The company reported that one API partner building AI tools for judges measured a 20 to 40 per cent improvement in downstream accuracy during an early evaluation. Another partner reportedly reduced its input-token use by approximately 51 per cent while finding all the source references it checked to be correct. These partner findings were reported by LexSelect and were not independently verified.
Developers can test the technology through LexStudio, the company’s self-serve platform for uploading documents and examining how their structure maps back to the original pages.