What data is included for this category?
Data for refinery financial reports comes from special disclosure content for the refining segments in listed refinery enterprises' periodic, semi-annual, and annual reports, plus monthly operation briefings released by industry associations. Updates follow two schedules: regular and ad-hoc. Regular reports are released quarterly, semi-annually, and annually. Ad-hoc announcements are issued alongside major events such as major unit maintenance shutdowns or product structure adjustments.
Document structure includes four modules: operating data overview, unit operation metrics, cost and expense breakdown, and profitability analysis. Core fields include crude oil processing volume, gasoline yield, unit manufacturing cost, and gross profit, with corresponding units of ten thousand tons, percentage, yuan/ton, and hundred million yuan. Content also includes structured tables and unstructured text analysis, with some reports embedding visual charts.
What constraints do these characteristics impose on the vector model and indexing workflow?
There are many structured fields and specialized terminology in refinery financial reports. Vector models must have semantic representation capabilities for industrial text to avoid misinterpreting specialized terms. Documents contain visual charts, so OCR text extraction must be completed before including content in the index. Otherwise, key data in charts cannot be retrieved.
The mixed regular and ad-hoc update rhythm requires the indexing system to support incremental synchronization, to avoid resource consumption and delays caused by full reindexing. Chunking must retain the association between fields and units, and cannot be truncated arbitrarily. Otherwise, retrieval results will fail to match complete business units.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Single chunk of refinery financial report content must cover complete unit operation metrics or cost breakdowns, to avoid truncating key data units |
chunk_overlap | 50–100 characters | Ensures critical cross-chunk terminology such as product yield is not split and lost, improving retrieval coherence |
embedding_model | Determined via actual testing (prioritize embedding models fine-tuned for industrial text) | Refining industry has a large number of specialized terms, requiring adaptation to semantic representation capabilities for industrial text |
embedding_api_url | Fill in according to the actual service address (fill in the API path of the corresponding port for local deployments) | Adapt to access requirements of different embedding services, ensuring vector generation requests can be sent correctly |
recall_top_k | 10–15 results | Refinery financial reports have many indicator dimensions, requiring recall of sufficient relevant chunks to cover multi-dimensional analysis needs |
similarity_threshold | 0.75–0.85 | Filter low-relevance non-industry term text, avoiding recall of irrelevant general financial report paragraphs |
incremental_sync_enabled | Enabled | Adapt to the regular updates and ad-hoc incremental release requirements of refinery financial reports, reducing resource consumption from full reindexing |
The parameter values provided on this page are common starting points for configuration setup. Actual values are affected by material format, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- Phenomenon: After switching the
embedding_model, the interface shows no synchronization progress, and it is not possible to revert to the original model. Cause: The original vector index cache was not cleared first. Switching the model directly causes conflicts in index generation tasks, and the system cannot recognize the vector format of the new model. - Phenomenon: A
404 page not founderror is returned when configuringembedding_api_urlto connect to a third-party embedding service. Cause: The API interface path of the vector model was not filled in correctly, or valid API key parameters were not carried in the request header. - Phenomenon: Financial report chart data embedded in the knowledge base cannot be retrieved by the vector index. Cause: The OCR text extraction configuration in the document parsing stage was not enabled, or a parsing plugin version matching the financial report format was not used.
How to Confirm Proper Configuration
- Upload a single refinery segment financial report snippet document, check the vector generation logs in the system console, confirm that the selected
embedding_modelhas loaded correctly and there are no abnormal errors. - Initiate a retrieval request containing refining industry specialized terminology, verify that the retrieved text chunks cover the target financial report fields and corresponding data.
- Upload a new ad-hoc announcement document, check whether the synchronization progress of the knowledge base index completes normally, confirm that the incremental update configuration has taken effect.
- Adjust the
recall_top_kparameter value, observe changes in the number of returned retrieval results, confirm that the configuration item has been synchronized to the retrieval workflow.
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.