What data for this category looks like
Data primarily comes from public annual reports, quarterly reports, and temporary announcements of domestic and overseas publicly traded white goods enterprises, plus monthly production and sales monitoring data released by industry associations. Updates follow disclosure timelines strictly: annual reports are updated once per year, quarterly reports are updated each quarter, and temporary announcements are released in real time alongside major business changes. Most documents are multi-page PDFs, containing structured tables such as category-specific revenue and raw material cost breakdowns, textual analysis paragraphs, and data charts. Core fields include category-specific revenue, core raw material procurement costs, offline and online channel revenue share percentages, and inventory turnover days. Common units are 100 million yuan, 10,000 units, and yuan.
What constraints do these characteristics impose on document parsing and chunking?
Multi-page PDF financial reports include structured tables and cross-page linked charts. The parsing process must retain the original row and column structure of tables and corresponding chart information to avoid damaging the corresponding relationship between data during chunking. There are many fine-grained business fields such as category-specific revenue and costs. Chunking must divide clear boundaries by business theme such as air conditioning sector revenue or upstream raw material costs to prevent cross-theme content from mixing into the same chunk and harming retrieval accuracy. Monthly production and sales data mostly exists in scattered Excel format, so structured parsing of table datasets must be supported. Differences in document length across different report cycles must be adapted to avoid overly long contexts exceeding model processing limits. Some historical financial reports are in scanned document format, so OCR recognition processes must be adapted to prevent loss of core field information due to recognition errors.
How to set the configurations
| Configuration Item | Recommended Value | Basis for This Value |
|---|---|---|
PARSE_OCR_ENABLE | Enabled | Some historical financial reports are scanned documents, requiring OCR to extract text and chart content |
CHUNK_SIZE | 800–1200 characters | Single chunk of financial report content must cover a complete single-category business segment to avoid splitting that disrupts business logic |
CHUNK_OVERLAP | 100–150 characters | Correlated content across themes must retain contextual cohesion to prevent loss of logical links after chunking |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Annual financial reports include multiple pages of charts and data tables, so single file volume is usually large |
TABLE_DATASET_PARSE_MODE | Map fields by column | Excel tables in white goods financial reports mostly organize production, sales, and revenue data by column, requiring accurate mapping of field corresponding relationships |
OCR_LANGUAGE | Chinese + English | Some financial reports of overseas-listed white goods enterprises include bilingual content, requiring accurate recognition of bilingual text |
The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Phenomenon: After the API is called to upload chart images attached to financial reports, the model cannot parse the revenue data within the images. Cause: OCR image recognition function for document parsing is not enabled, or legal format and size limits for image uploads are not configured.
- Phenomenon: When a single quarterly financial report PDF is parsed, the chunking results include irrelevant content from adjacent documents, and contextual isolation for a single document cannot be achieved. Cause: Trigger rules for chunking by single document are not set, or the
CHUNK_OVERLAPvalue exceeds a reasonable range, causing cross-document content to be linked. - Phenomenon: After Excel-format production and sales data is imported, the column corresponding relationship between category-specific sales volume and revenue is lost, and target fields cannot be accurately retrieved. Cause:
TABLE_DATASET_PARSE_MODEis not configured to map fields by column, or the matching logic between table headers and business fields is not verified.
How to confirm configurations are properly set
- When a single scanned financial report PDF is uploaded, check the parsed text content, confirm there is no obvious garbled code, and verify the recognition accuracy of core fields.
- When a multi-page structured financial report is uploaded, check the chunking results, confirm each chunk contains complete business theme content, and no cross-theme content mixing occurs.
- When an Excel-format production and sales data table is imported, check the field mapping relationship of the dataset, and confirm the corresponding logic between column titles and business fields is correct.
- When chart images in financial report format are uploaded via API, check the parsed text content, confirm that revenue, sales volume and other data within the images are correctly extracted.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.