How Tencent WorkBuddy’s TextIn xParse Turns PDFs into Structured Data for LLMs

TextIn xParse: Not OCR, but Intelligent Document Infrastructure

Faced with the pain point of converting a PDF version of Tao Zhexuan’s ICM 2026 talk into an editable Word document, the author discovered that traditional OCR‑based conversion loses the document’s skeleton—tables become plain lists, page breaks break content, and paragraph order gets scrambled.

TextIn xParse, a general‑purpose document‑parsing service built on more than 20 years of multimodal text‑intelligence research, can transform PDFs (including scanned files), Word, PPT, and images into structured Markdown/JSON while preserving original hierarchy, table logic, image‑text relationships, and cross‑page content.

Real‑World Test: From Unreadable to Precise Positioning, a Complete Workflow

The test document is a 52‑page PDF of Tao Zhexuan’s ICM 2026 speech, featuring mixed graphics, cross‑page paragraphs, headers/footers, and slide‑style heading structures that typically challenge AI agents.

Step 1: Add the WorkBuddy Connector with One Click

In WorkBuddy’s sidebar, the user opens the “Connectors” panel, locates the TextIn xParse card, and clicks the “+” button. No adaptation code is required, and the TextIn account is authorized through a simple linking step.

Step 2: Parse the PDF and See the Difference Immediately

Inside a WorkBuddy chat window the user issues the command:

这是陶哲轩《Mathematics in the age of AI》演讲PDF，转成可编辑的word，保留了完整原版结构，不限于大纲、图表、页眉、层级与脚注标记等信息

The 52‑page PDF is processed in seconds, producing a DOCX that faithfully restores title hierarchy, correctly matches images to captions, and retains the full structural layout.

Step 3: Feed the Structured Data to an LLM for Real Work

With the clean Markdown output, the author prompts an LLM to extract the speech’s core arguments and key citations. The LLM returns answers that quote the original wording (including years and DOI numbers) and accurately locate each point within the document’s hierarchy—something the raw text input could not achieve.

Summary and Evaluation

The combination of TextIn xParse and WorkBuddy solves the “first‑kilometer” problem for AI agents handling real‑world documents: without high‑precision structured parsing, downstream retrieval‑augmented generation, summarisation, or Q&A would be built on a shaky foundation.

For individual developers and small‑to‑medium businesses, the service offers a generous free quota of 1,000 pages per day, lowering the barrier to entry.