11,044 documents refined into 116 processes, 0 fabricated

The conversion rate isn't impressive, but every piece of data can withstand scrutiny.

11,044 documents from the phosphorus chemical industry (GB/T standards, patents, papers) were fed into the pipeline, resulting in: 116 real processes + 279 BOM data entries, 0 errors.

An 86 / 11,044 conversion rate is hard to look at. But we care more about another number: 10,958 documents were honestly skipped.

The Zero Data Fabrication Rule

The LLM in the pipeline does only one thing: judge whether a document contains real process information. If yes, extract it verbatim; if no, skip it. It is not allowed to 'complete based on industry knowledge'.

Why? Because in industrial settings, a single fabricated process parameter entering the production system is an incident. The model saying 'I recall wet-process phosphoric acid is typically at 80°C'—that's common sense in a paper, but a potential hazard next to a reactor.

What does the stored data look like (real samples)

• Wet-process phosphoric acid graded utilization: Equipment=Scrubber, Temperature 50~60°C, NH₄H₂PO₄ concentration 0.5~1% (Source: Patent CN100441502C)
style="padding-left:20px;margin:12px 0;">
• Thermal process phosphoric acid production: Electric furnace/Reaction tower, above 1450°C, concentration 80~85%
style="padding-left:20px;margin:12px 0;">
• Wet-process phosphoric acid decomposition of phosphate rock: Reactor/Agitation/Washing, 70°C, P₂O₅ 30%, duration 3.0h

Each entry can be traced back to its source document. What's extracted is not 'AI-interpreted knowledge', but data that was originally present in the documents.

The complete pipeline chain

Collection (11,044 documents stored, real URL+abstract) → LLM structuring (86 documents successfully converted) → Process/BOM entry into pskb knowledge base → Assetization (into data_assets, independently valuatable) → Consumption by digital employees (direct query by process optimization engineers, quality inspectors).

Full batch validation date: 2026-08-02, validation report available in docs/.

BossAgents: The first step in industrial data assetization is to make data 'real' first, then talk about its 'value'.

All data in this article can be traced in the BossAgents technical documentation repository · BossAgents · Private AI Agents for the Boss