Zhihe GrowthA PDF's visual layout can differ from its internal text order. A reader sees a model, voltage and tank size in one row, while extraction may return all tank values before all model names. Pairing adjacent text fragments can then invent relationships absent from the source.
Input and comparison boundaries
For reproduction, inspect the two PDF pages before plain text and table arrays. Verify model-row pairing and second-page restrictions. If a person corrects extraction, retain the raw value, correction, coordinates and reason rather than only a clean JSON file. OCR confidence can prioritize review but is not the probability a specification is correct. This experiment uses a digitally generated PDF and neither runs scanned-document OCR nor ranks OCR providers.
The synthetic manual contains two pages and two fictional models. Page one covers supply and tank size; page two covers restrictions. The fixture uses a valid column-major writing order while displaying a normal table. It is owned teaching material, not a customer manual or representative sample of all PDF layouts.
We execute pypdf's ordinary extract_text and pdfplumber's extract_table. These perform different tasks. Absence of table structure in a plain-text output is not a comparative library accuracy score. The example demonstrates why a downstream system must understand its input representation.
Inspect the actual outputs
Page-one plain text begins with Tank, 20 L and 10 L, followed by Model, AX-220 and AX-110, then Supply, 220 V and 110 V. Row relationships must be recovered rather than inferred from textual adjacency.
| Source row | Supply | Tank |
|---|---|---|
| AX-220 | 220 V | 20 L |
| AX-110 | 110 V | 10 L |
document-results.json preserves both outputs. On this clearly ruled fixture, pdfplumber recovers both tables exactly as authored. That means these two pages pass; it does not mean PDF extraction is universally 100% accurate. Scans, borderless layouts and merged cells are outside this executed example.
Retain headers, location and provenance
A record should contain the original hash, physical page, table identifier, row entity, column heading, value, unit, footnote and coordinates. A converted CSV alone is insufficient when a reviewer needs to return to the original. Repeated headings on a continuation page must be distinguished from genuine data rows.
| Layout issue | Review action | Unsafe assumption |
|---|---|---|
| Empty cell | Determine missing value or merged structure | Inherit the row above |
| Footnote marker | Preserve marker-to-note association | Retain only the number |
| Multiple columns | Reconstruct region order | Concatenate the whole page |
| Scanned text | Keep image and OCR evidence | Treat recognition as an original fact |
| Repeated heading | Mark table continuation | Create a product record |
Use field-level acceptance
For each critical field, record the expected value, extracted value, unit, entity, conditions and error type. A correct unit assigned to the wrong model fails. A correct number without its limiting footnote is incomplete. High character-recognition accuracy cannot compensate for a small number of dangerous parameter errors.
Classify omissions, transcription errors, row or column shifts, lost units, lost conditions and version mixing separately. Sample difficult layouts deliberately, including merged cells, page boundaries and nearly identical model names. A single aggregate percentage can hide the defects that matter.
Know when human review is required
Safety, certification, warranty and quotation fields need review before entering the fact register. Disagreement between parsers is not resolved reliably through majority voting, since tools may share an OCR mistake. Reviewers need the original page, not merely competing extracted text.
The pypdf documentation explains extraction and OCR boundaries. pdfplumber's reference documents table and coordinate interfaces. media_labs.py reproduces this example with package versions recorded. Both rendered pages remain available for visual inspection.
Zhihe Growth's document-ingestion standard
For spanning tables, check continued column definitions, repeated headers and footnote scope. Both fixture pages contain complete headers; complex cross-page merging was not tested. Success here does not establish reliability on arbitrary scanned manuals.
For exporters working with Zhihe Growth, ingestion should preserve model, unit and operating scope rather than stop at bulk PDF-to-text conversion. Customers can sample high-risk fields and follow each from the original page through the fact record to Chinese and English website content.
Continue with unit validation, evidence locations and SSOT mapping. The technical FAQ offers short answers; the advanced package contains original fixtures and output. Correct ingestion is an engineering result, not evidence of improved external AI citation rates.