What the data for this category looks like
Joint-stock bank financial report data primarily comes from officially disclosed annual, semi-annual, and quarterly reports. Quarterly reports are released within one month following the end of each quarter. Annual reports are released within four months following the end of the fiscal year. Document structures include consolidated balance sheets, consolidated income statements, consolidated cash flow statements, and detailed business and financial notes. Core fields cover total assets, non-performing loan balances, capital adequacy ratios, provision coverage ratios, and other relevant metrics. Most metrics use RMB 100 million as the measurement unit, while some regulatory indicators use percentage formatting.
Constraints imposed on document parsing and chunking
The large volume, fixed field format, and high-frequency updates of joint-stock bank financial reports create multiple constraints for the document parsing and chunking process. A single annual report can span dozens of pages. Streaming parsing must be supported to avoid memory overflow. Fields are tightly bound to their units. For example, total assets uses 100 million RMB as its unit, while non-performing loan ratios use percentage notation. Chunking must retain semantic associations between adjacent fields, and avoid splitting explanatory content for the same metric that spans multiple pages. Regulatory indicators have fixed disclosure positions. Splitting cross-page metric definitions and numerical values into separate chunks must be avoided. High-frequency updated quarterly and annual reports require version verification to prevent repeated parsing of old version data.
How to set the configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
PARSE_CHUNK_SIZE | 800–1200 characters | Adapts to the length of core metric groups in joint-stock bank financial reports, avoids splitting complete financial data chunks |
PARSE_CHUNK_OVERLAP | 150–200 characters | Retains cross-chunk semantic associations for metrics, avoids splitting long explanatory paragraphs in financial notes |
MAX_DOC_PARSE_SIZE | 500 MB | Adapts to the maximum file size of a single annual financial report, prevents parsing failures for large documents |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Provides sufficient time for streaming parsing and format extraction of long documents |
ENABLE_TABLE_PARSE | Enabled | Financial reports contain large numbers of structured financial tables, must retain the row and column semantic structure of tables |
VERSION_CHECK_ENABLE | Enabled | Matches the latest disclosed financial report version, prevents repeated parsing of old version data |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require targeted analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Symptom: After uploading a publicly available financial report link, the parsing result only contains the page title with no body content. Cause: PDF parsing adaptation for
ENABLE_LINK_PARSEis not configured. Link parsing added in v4.8.13 requires additional configuration of the document type whitelist. - Symptom: Context references after chunking show financial metrics separated from their units. Cause:
PARSE_CHUNK_OVERLAPis set too low, failing to retain semantic associations between adjacent metrics, leading to cross-chunk metric explanations being split. - Symptom: Fields are missing after uploading documents split by a local vector model to the server. Cause:
PARSE_CHUNK_SIZEand local chunking parameters are not unified. Different chunking logic damages semantic blocks.
How to confirm configurations are correctly applied
- Upload a single-quarter financial report PDF, view the parsed chunk list, and confirm each chunk contains complete financial metrics and explanatory content.
- Call the document parsing API, check if the returned
chunk_textfield retains complete row and column information for tables, with no formatting errors. - Upload both old and new versions of the same financial report, confirm the parsing result only retains chunk data from the latest version, with no duplicate entries.
- Adjust the
PARSE_CHUNK_SIZEparameter, re-parse the same document, and compare changes in chunk results to confirm the configuration takes effect.
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.