Why Unstructured Data—80% of Enterprise Knowledge—Blocks AI Adoption

A widely underestimated fact is that more than 80% of an enterprise’s knowledge is stored in unstructured forms such as contracts, invoices, reports, drawings, and process documents.

When building Retrieval‑Augmented Generation (RAG), knowledge bases, or AI agents, the first obstacle is not retrieval or generation but converting the knowledge trapped in these documents into structured information that AI can understand, index, and invoke. Poor handling of this step leads to “garbage in, garbage out” for all subsequent stages.

In practice, the difficulty is far greater than most anticipate. Complex layouts—cross‑page tables, nested tables, mixed text‑image flows, engineering drawings—currently achieve only 65‑70% extraction accuracy. Table extraction is dubbed the “last frontier” because tables hide critical business data such as amounts, dates, clauses, and metrics. Multimodal fusion is incomplete; text and images are processed separately, causing loss of key visual information. General‑purpose parsing models struggle with vertical documents (financial contracts, medical records, manufacturing drawings), and even basic OCR still contains gaps. Unstable parsing quality fed directly into RAG or knowledge bases becomes the first domino that can topple an AI agent deployment.

This is not a problem that can be solved by simply calling an API; it requires deep engineering effort. Key questions include: how to chunk documents—should tables be kept whole or split by rows? How to bind images with surrounding text? How to segment long documents? How to measure parsing quality and continuously improve it? Each decision involves trade‑offs.

To address these challenges, the DACon 2026 Beijing forum dedicated a session to “Unstructured Data Processing and Knowledge Acquisition.” The session focuses on the gritty, labor‑intensive engineering work needed to release knowledge from documents into AI‑readable, searchable structures.

The forum agenda covers:

Deep‑water document parsing: high‑precision extraction of complex, cross‑page, nested tables and why tables are the “last frontier.”

Multimodal content understanding: joint interpretation of mixed text‑image layouts and alignment of visual and textual information.

Domain adaptation: tailoring generic parsing models to vertical documents such as financial contracts, medical records, and manufacturing drawings.

Chunking strategies: engineering practices for whole‑table vs. row‑level splitting, image‑text binding, and long‑document segmentation.

Industry scenario practice: real‑world deployments in finance, manufacturing, and healthcare.

Parsing quality evaluation: methods to measure accuracy and strategies for continuous improvement.

Each topic follows a “problem → method → architecture → case → metrics → pitfalls” structure, avoiding pure product pitches and emphasizing concrete engineering experiences.