What the data for this category looks like
Data for thermal energy due diligence reports primarily comes from monthly operation reports of thermal energy supply enterprises, thermal pipe network monitoring data from government regulatory authorities, user payment ledgers, and pipe network operation and maintenance logs. The document update cycle centers on the annual heating season, supplemented by monthly operation summaries and quarterly operation and maintenance reviews. Document structures include modules such as pipe network parameter tables, heat supply statistics, revenue and cost details, user coverage scope, and operation and maintenance fault records. Fields include numerical items with clear units: heat supply (gigajoules), pipe network pressure (megapascals), payment amount (yuan), operation and maintenance duration (hours). Semi-structured operation and maintenance log paragraphs and a small number of pipe network layout diagrams are also included.
What constraints do these characteristics impose on the document parsing and chunking link
The mixed format of thermal energy due diligence reports requires precise distinction between structured tables, semi-structured logs, and visual elements during parsing. This prevents merging different types of content into the same chunk. Fixed-cycle updated unified templates require chunking to retain the binding relationship between fields and their corresponding values. Field names and values cannot be split apart. Numerical fields with clear units require retaining the association between numerical values and units during parsing and chunking. This prevents situations where numerical values lack matching units after splitting. Long tables and cross-page operation parameter tables require that continuous rows of the same table are not split across different chunks. This ensures the semantic integrity of the table.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_TABLE_MODE | "mixed" | Adapts to the mixed table format of structured reports and semi-structured operation logs in thermal energy due diligence reports, and fully retains the semantic structure of tables |
CHUNK_SIZE | 800–1200 characters | Thermal energy due diligence reports have numerous fields with close associations. This range balances the integrity of single-block information and subsequent recall accuracy |
CHUNK_OVERLAP | 150–200 characters | Avoids splitting associated content across fields, and ensures contextual continuity between chunks |
MAX_TABLE_ROWS_PER_CHUNK | 10–15 rows | Thermal pipe network parameter tables typically contain multiple consecutive rows of data. Limiting the number of table rows per chunk prevents chunk overload |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Large thermal energy due diligence reports include multi-page tables and logs. This duration covers the complete parsing process |
ENABLE_GPU_PARSE | true | Adapts to complex format parsing requirements. GPU acceleration must be enabled to process multi-module mixed thermal reports, and this setting also adapts to the calling logic of MinerU-like parsing services |
The parameter values provided on this page are common 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. It is recommended to test on your own samples before finalizing the configuration.
Three common errors
- Phenomenon: Tables in some thermal energy due diligence reports cannot be chunked correctly, and table fields appear empty in returned results. Cause: The
PARSE_TABLE_MODEconfiguration is not set to the mixed format adaptation mode. Only plain text parsing is enabled, and the table structure is ignored. - Phenomenon: An out-of-video-memory error occurs after starting the parsing service, and logs show no available GPU is detected. Cause: The GPU driver 570 and CUDA 12.8 environment are not mounted during Docker deployment, and the
ENABLE_GPU_PARSEconfiguration item is not enabled. This causes the parsing service to only use CPU resources and unable to process the complex format of large thermal energy reports. - Phenomenon: Operation and maintenance images in imported thermal energy due diligence reports do not carry a preset domain name, and cannot be loaded normally during conversations. Cause: The image external link binding configuration during document parsing is not enabled. Only local image paths are extracted, and no accessible domain name prefix is replaced.
How to confirm the configuration is correct
- Upload a standard sample of a thermal energy due diligence report, and check the parsed chunk results. Confirm that table content is complete and not split into unrelated chunks.
- Check the service running logs, and confirm that the GPU device is properly identified. No error messages about insufficient video memory or unmounted devices appear.
- Check the parsed text, and confirm that numerical values and their corresponding units are bound in the same chunk. No splitting occurs.
- Import a thermal report containing operation and maintenance images, and confirm that the image external links have added the preset domain name. Images can be loaded normally in the conversation interface.
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.