跪拜 Guibai
← All articles
AIGC

TextIn xParse Turns a Mess of Supplier Quotes into a Traceable Procurement Report

By 一只牛博 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Procurement teams waste hours manually flattening merged cells and copying revision pages before they can even start comparing bids. A parsing layer that preserves table structure and evidence location shifts the bottleneck from document triage to actual decision-making, and the reusable Agent prompt means the rules are encoded once and applied consistently across future rounds.

Summary

Supplier comparison for a bridge-monitoring project starts with four vendors sending incompatible formats: multi-level header Excel, Word with tables inside tables, scanned PDFs, and package-discount sheets. TextIn xParse, connected through WorkBuddy, restores table hierarchies, merged cells, and parent-child relationships before any analysis begins. Without it, a standard model misreads nested Word tables as missing line items. With xParse, the same model correctly maps equipment codes, flags a hard negative deviation on a 0.8Hz frequency lower limit, and isolates optional accessories so they don't inflate the effective mandatory quote.

The workflow produces an 18-page Word report that ranks suppliers conditionally — not by lowest price — and lists seven cross-referencing anomalies alongside their source page or worksheet. A reusable Agent prompt encodes the entire procurement caliber: mandatory vs. optional, discount handling, technical thresholds, and a rule that missing data stays marked 'Not Provided' rather than guessed. The connector's free tier covers 1,000 pages per day; a baseline test on a clean Excel showed a 1.7× speedup in full response time compared to plain parsing.

Takeaways
TextIn xParse restored 5 major systems and 12 equipment line items with parent-child relationships from a multi-sheet Excel, including blank rows that belonged to repeated parent categories.
Without xParse, a Hy3 model misjudged PK-D and PK-E as missing quotations in a Word document; the data was present but nested in a separate table layer.
After enabling xParse, the same model correctly identified Supplier B's B-01 frequency range of 0.8–120Hz as a hard negative deviation against the procurement lower-limit requirement.
Supplier D's PDF contained a deliberately altered A-02 quantity of 2 with unchanged unit price and amount; xParse preserved the anomaly for manual cross-referencing review.
A baseline test on a clean Excel showed a full-response time of 2 minutes 8 seconds with xParse versus 3 minutes 39 seconds without — a 1.7× speedup in one recorded run.
The final output is an 18-page, 5,955-word editable Word report with supplier summary, technical deviations, business risks, recommendation order, and items requiring manual confirmation.
A reusable System Prompt encodes the entire procurement caliber: mandatory vs. optional items, package discounts, cross-referencing rules, and a strict 'mark as Not Provided, don't guess' policy.
The connector's free tier provides 1,000 pages per day; parsing time for a single file was 328ms, while the full WorkBuddy response includes subsequent reading and generation.
Conclusions

The core value isn't speed — it's that xParse prevents a model from hallucinating missing line items when data is merely nested in a different table layer, a failure mode that silently corrupts procurement decisions.

Encoding procurement rules into a reusable Agent prompt turns a one-off demo into a repeatable process; the real leverage is that future rounds don't need to re-explain what counts as mandatory, optional, or a discount.

The deliberate A-02 quantity anomaly test reveals a design philosophy worth noting: the system flags discrepancies for human review rather than auto-correcting, which is the right posture for financial decisions where a wrong override carries real cost.

1.7× speedup on a clean Excel is a nice headline number, but the more consequential metric is that nested Word tables and scanned PDFs become machine-readable at all — without xParse, those formats produce broken comparisons.

Concepts & terms
TextIn xParse
A document-parsing connector within WorkBuddy that restores table hierarchies, merged cells, nested tables, and parent-child relationships from Excel, Word, and PDF files, outputting structured Markdown and JSON for downstream analysis.
WorkBuddy
An AI assistant platform that supports connectors like TextIn xParse for document analysis and cross-file comparison, allowing users to build reusable Agents with custom system prompts.
Cross-referencing (勾稽)
In procurement, the process of verifying that quantity × unit price matches the line-item amount, that tax calculations are correct, and that totals across different table layers are consistent. Anomalies are flagged for manual review rather than auto-corrected.
Procurement normalization
Mapping different suppliers' field names, table structures, and pricing calibers (mandatory vs. optional, package discounts, bundled accessories) onto a single comparison framework keyed by the buyer's equipment codes.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗