What the data for this category looks like
Financial leasing financial statement data mainly comes from project ledger systems, rent collection modules, leased asset valuation archives, and regulatory submission reports. Data updates follow a quarterly basic cycle. Annual financial statements require full project reviews. Each financial statement document includes fields such as project number, lessee entity information, original value of leased assets, receivable rent for each period, overdue amount, and asset depreciation details. Units include RMB yuan, number of leased assets (sets/units), lease term in months, and some documents include structured extracted content from paper scans.
What constraints do these characteristics impose on the vector models and indexing link
The multi-source mixed data characteristics of financial leasing financial statements require vector models to support mixed embedding of structured fields and unstructured text. Quarterly high-frequency updates and annual full project review requirements require the indexing system to support incremental updates and fast full reconstruction. Fields include multiple types of units such as amount, quantity, and lease term. Field standardization mapping must be completed in advance to avoid unit differences interfering with semantic similarity calculation during embedding. Each document covers multiple leasing projects. Segmentation must split content by project entity to avoid cross-project semantic confusion affecting recall accuracy.
Configuration settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
EMBEDDING_MODEL | text-embedding-3-large | Adapts to the mixed embedding requirements of structured numerical values and multi-type semantic text in financial statements, and provides more stable semantic understanding of multi-unit fields |
CHUNK_SIZE | 800–1200 characters | Each financial statement includes multiple leasing projects. After splitting by project, this length adapts to the model context window and avoids cross-project semantic mixing |
INDEX_INCREMENTAL_UPDATE | Enabled | Matches quarterly high-frequency update requirements, reducing time and computing costs of full reconstruction |
FIELD_NORMALIZATION_RULE | Unified mapping by currency and unit | Eliminates embedding interference from multi-unit fields in financial statements, improving the accuracy of semantic similarity calculation |
RECALL_TOP_K | Top 10 entries | Covers the number of related projects required for financial statement analysis, avoiding excessive recall leading to context redundancy |
PARSE_STRUCTURED_FIELD | Enabled | Combines structured numerical fields and semantic text for embedding, improving the matching degree between recall results and financial statement analysis requirements |
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 specific analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Issue: Swapping
EMBEDDING_MODELleads to abnormal recall results in the existing knowledge base. Cause: The existing index uses vectors generated by the old model, and vector embedding results are not updated synchronously, leading to inconsistent vector spaces between old and new vectors. - Issue: Trying to select an embedding model other than
text-embedding-ada-002returns the errorundefined model must match "^(text. Cause: API permissions for the corresponding model are not enabled, or the full model name prefix is not filled in the configuration item. - Issue: Knowledge base recall results are not sorted according to semantic similarity. Cause: Field standardization rules are not enabled, and numerical differences in multi-unit fields interfere with similarity calculation, leading to abnormal sorting logic.
How to confirm the configuration is correctly set
- Upload a single test financial statement document, and check the vector embedding log to confirm that structured fields and non-text content are correctly embedded.
- Perform an incremental update operation to verify that the index only updates newly added or modified documents, and does not trigger a full reconstruction process.
- Initiate a financial statement keyword search, and check that the sorting logic of recall results matches semantic similarity matching.
- After replacing the embedding model, trigger a full index reconstruction to verify that there is no abnormal fluctuation in the recall results of the existing knowledge base.
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.