Knowledge Base Retrieval and Recall for Joint-Stock Bank Financial Report Analysis

Joint-stock bank financial report data is sourced from internal financial accounting systems, official disclosure platforms, and regulatory reporting

What the data for this category looks like

Joint-stock bank financial report data is sourced from internal financial accounting systems, official disclosure platforms, and regulatory reporting databases. Update cadence varies by report type: annual reports are released within four months after the end of the fiscal year, semi-annual reports within two months after the end of the half-year, quarterly reports within one month after the end of the quarter, and temporary operating announcements are updated as needed. The structure of a single report document includes three parts: financial statements, regulatory compliance indicators, and business operation analysis. Core fields include total operating revenue, net profit, core tier 1 capital net amount, non-performing loan balance, etc. Units are uniformly billion yuan or ten thousand yuan.

What constraints these characteristics impose on knowledge base retrieval and recall

Multi-source data sources require format alignment rules to avoid format conflicts for the same indicator across different channels, which would reduce retrieval accuracy. High-frequency update cadence requires incremental synchronization mechanisms to reduce resource consumption from full synchronization while ensuring data timeliness. Long document structures require adjusted segmentation rules to preserve contextual connections between report sections and avoid breaking the complete logic of financial indicators. The specificity of professional fields requires configuration of precise field filtering and synonym expansion to ensure retrieval results match the core needs of business analysis.

How to set configurations

Configuration ItemRecommended ValueRationale
chunk_size800–1200 charactersFinancial data segments of joint-stock bank reports are closely linked. Excessive length will lose contextual connections, while insufficient length will break the logical integrity of report sections
recall_top_kTop 10–15 resultsFinancial report analysis needs to cover multi-dimensional business indicators and regulatory data. Too few results cannot support complete analysis, while too many will introduce irrelevant retrieval results
similarity_threshold0.75–0.85Financial report professional terms have high recognition. A threshold that is too low will introduce irrelevant document fragments, while a threshold that is too high will miss core content with precise matches
sync_intervalEvery 6 hoursJoint-stock bank temporary announcements have a relatively high update frequency. Incremental synchronization can ensure timeliness while reducing server load
parse_field_filterOnly retain operating revenue, net profit, core tier 1 capital net amount, non-performing loan balanceFilter redundant non-core field content to improve the targeting and relevance of retrieval results
vector_dim1536Matches the output dimension of general embedding models to ensure compatibility of vector storage and retrieval

The parameter values provided on this page are all common recommended starting points 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.

Three common errors

  • When the /api/v1/kb/search interface is invoked, it returns {"code":514,"statusText":"invalid kb id"}. The cause is that the kb_id parameter of the knowledge base is not configured correctly, or the passed kb_id does not match the actually created knowledge base ID.
  • After field filtering rules are configured, retrieval results still include non-financial report documents. The cause is that the field filtering function is not enabled in the document parsing configuration, or the configuration format of the filtering rules is incorrect.
  • Retrieval response times out when PGVector is used for knowledge base storage. The cause is that index parameters are not adjusted based on data volume, and the caching mechanism of the vector index is not enabled, resulting in insufficient single-table query efficiency.

How to confirm configurations are properly set

  • Invoke the knowledge base search test interface, pass a query term related to core financial report indicators, and verify that the document type of returned results matches the target financial report data.
  • Synchronization task logs are reviewed to confirm that the incremental synchronization task executes automatically according to the configured cycle, with no failure alarm records.
  • Vector storage index configurations are checked to confirm alignment with the current knowledge base’s data scale, and that retrieval response times meet business requirements.
  • The field list of a single retrieval result is exported, and verification is performed to confirm only preset core business indicator fields are included.

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.