What the Data for This Category Looks Like
Securities financial report data comes from official disclosure platforms of the Shanghai, Shenzhen, and Beijing Stock Exchanges, and designated information disclosure channels of listed companies. Update rhythms fall into two categories: regular and temporary. Regular disclosures include quarterly, semi-annual, and annual reports, released during fixed disclosure windows. Temporary disclosures include performance forecasts, correction announcements, and other materials, with no fixed release schedule. Document structures combine structured financial tables and long explanatory text, covering sections such as consolidated financial statements, management discussion and analysis, and financial notes. Fields include net profit attributable to parent, non-recurring net profit, net cash flow from operating activities, and others. Units are mostly RMB yuan, ten thousand yuan, or hundred million yuan.
Constraints on Vector Models and Indexing
Mixed regular and temporary update rhythms require the indexing system to support both incremental update and full reindexing modes, to meet synchronization needs for different disclosure scenarios. Single financial report documents vary significantly in length—annual reports can reach tens of thousands of characters. This requires designing reasonable text segmentation rules to avoid semantic cutting errors. The mixed structure of structured tables and unstructured text requires vector models to adapt to encoding logic for both structured semantics and unstructured text. The standardized characteristics of fields and units require retaining metadata tags during indexing, to provide a basis for subsequent precise recall.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
embedding_model_normalization | Enable as required by the target model; enable for models that do not support normalization | Adapt to the vector output format of different embedding models, avoid deviation in similarity calculation |
chunk_size | 800–1200 characters | Adapt to the text length of single financial reports, avoid overly fragmented or overly long semantic cuts |
recall_top_k | Top 8–12 results | Cover key information across multiple chapters of financial reports, avoid insufficient recall or redundancy |
max_context_token | 16000–32000 tokens | Match the total text length of securities financial reports, ensure recalled content is fully passed to the large model |
vector_index_refresh_interval | 1 hour (during regular report disclosure windows) / 24 hours (normal periods) | Adapt to the update rhythm of financial reports, balance synchronization timeliness and system load |
force_rebuild_index | Trigger as prompted by the interface | Resolve index rebuilding issues after switching vector models, force reconstruction of indexes compatible with the new model |
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 case-by-case analysis. Testing on local samples is recommended before finalizing settings.
Three Common Configuration Mistakes
- Phenomenon: After switching the knowledge base vector model, the index progress does not update and rollback to the original model is not possible. Cause: The
force_rebuild_indexconfiguration was not triggered. The system did not rebuild the index to adapt to the new model, resulting in index incompatibility with the current model. - Phenomenon: The context token passed during knowledge base query exceeds the preset threshold, leading to increased query latency. Cause: The
max_context_tokenconfiguration was not adjusted based on the average length of single financial reports, and a value exceeding system load was set directly. - Phenomenon: Charts and screenshots attached to financial reports cannot be indexed and updated correctly. Cause: The version is not upgraded to 4.8.23 or above. Older versions do not support incremental update functionality for image vector indexing.
How to Verify Configuration Completion
- Access the vector model configuration page, confirm that the value of
embedding_model_normalizationmatches the currently used embedding model. - Upload a single quarterly financial report, check that the parsed text segmentation length falls within the preset
chunk_sizerange. - Submit an incremental indexing task, confirm that the task progress bar updates normally and there are no failure prompts.
- Initiate a simulated query, verify that the recall results include core fields of the financial report and corresponding unit information.
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.