LlamaIndex enhances the reading of complex tables and charts with a 4-billion-parameter model

OCRia-1 improves table, chart, and formula extraction from PDFs. The LlamaIndex model achieves 87.4 on ParseBench.

A poorly transcribed financial table can shift a value into the wrong column without producing any obvious error. In an insurance contract, a coverage limit can end up associated with the wrong category. These discrepancies become particularly problematic when the documents are subsequently handed over to artificial intelligence agents, which use the extracted information to perform searches, produce analyses, or prepare decisions. LlamaIndex seeks to reduce these errors with OCRia-1, a four-billion-parameter vision and language model specialized in reading complex documents. Integrated into LlamaParse, it achieves a score of 87.4 on ParseBench, ahead of the other parsing models compared in the study, while offering optimizations designed for processing large document volumes. Reading a document is not just about recognizing its characters

Optical Character Recognition, or OCR, transforms text present in an image or digitized document into actionable content. But correctly recognizing words is not enough to reconstruct the structure of a page. A PDF can contain multiple columns, nested tables, annotations, equations, or charts whose meaning depends directly on their layout.

Digitally created documents do not always solve this problem. Their text layer allows for the retrieval of characters and their positions, without necessarily indicating which elements form a title, a paragraph, or a table header. Scanned documents add difficulties related to resolution, print quality, and the absence of directly accessible text.

OCRia-1 processes this information from a visual representation of the page. It primarily produces Markdown, while using HTML to preserve tabular structures and LaTeX to transcribe mathematical expressions. A table with multiple header levels, for example, must retain the relationships between each label and the corresponding values, rather than becoming a simple succession of text lines.

This distinction interests companies that use document retrieval systems or agents capable of querying archives. An error introduced at the time of extraction can subsequently be carried over into a response, a calculation, or a report, even if the model responsible for the analysis functions correctly. Table reconstruction at the core of training

Tables constitute one of the primary development focuses of OCRia-1. The model is trained to recognize merged cells, multi-line headers, irregular grids, and sets containing several thousand cells. These structures are common in accounting documents, regulatory reports, statistical appendices, and insurance files.

On the table component of ParseBench, OCRia-1 reaches 91.5 points, compared to 89.2 for Chandra-OCR-2, its closest competitor among the compared systems. On the tables category of olmOCR-bench, it scores 92.9 points, ahead of Chandra-OCR-2 at 92.1.

The model also takes the lead on PulseBench-Tab, which comprises 1,820 tables in nine languages, with 84.1 points against 82.4 for the following models. On the long table categories of the French benchmark fr-bench-pdf2md, it achieves 83.2 and 81.9 points, respectively.

These evaluations do not solely measure the presence of the correct characters. They also seek to determine whether the cells have been placed back into the correct structure. This requirement is essential when a document must be processed automatically: an exact value, but assigned to the wrong column, remains incorrect information. Charts transformed into structured data

Charts present a different problem. Their content is not necessarily written in text form: one must interpret the height of a bar, follow a curve, identify axes, and associate data with the correct series. Some charts do not even feature numerical values directly inscribed next to data points.

OCRia-1 seeks to reconstruct this information in the form of HTML tables. Rather than producing only a general description of a chart, the model extracts data points and associates them with the rows and columns necessary for their interpretation.

On the ParseBench chart evaluation, it achieves 85.1 points, compared to 66.1 for LightOnOCR-3-4B and 65.1 for Chandra-OCR-2. The 19-point gap with the second-place model represents one of the most pronounced results of the study.

This capability applies to economic reports, scientific publications, and professional documents in which numerical information is presented exclusively in visual form. The model can prepare this data for subsequent analysis, but its extraction remains an estimation depending on the legibility of the chart, its scales, and its complexity. The benchmark results do not guarantee an exact rendition of every value. Formulas, superscripts, and annotations no longer systematically disappear

The formatting of a document can carry information just as important as the text itself. A superscript in a formula, a footnote subscript, a struck-through passage in a contract, or a checked box in a form can alter the meaning of the extracted content.

OCRia-1 is trained to preserve these elements. It uses appropriate tags to retain superscripts, subscripts, underlined, or struck-through passages, and transcribes mathematical expressions into LaTeX. It also seeks to respect the reading order of a multi-column page.

On the formatting dimension of ParseBench, the model scores 83.4 points, compared to 67.9 for Surya OCR 2, which holds second place. For the text itself, the results are less differentiated: OCRia-1 reaches about 90 points, a level comparable to that of several competitors.

The announced gains are therefore concentrated on the information that transcription tools have historically struggled the most to preserve, rather than on the simple recognition of ordinary paragraphs. A specialized Qwen3.5 model with 100,979 training pages

The technical report specifies that OCRia-1 relies on Qwen3.5-4B, adapted for document transcription through targeted training. Researchers assembled a corpus of 100,979 pages, consisting of 82,943