What this type of data looks like
Financial leasing investment research data mainly comes from lease project contracts, rent calculation sheets, corporate credit reports, industry regulatory documents, and third-party due diligence reports. The data update rhythm adjusts with project progress. Full-cycle data for a single project from project initiation to launch is synced throughout the project lifecycle. Daily industry research reports and regulatory documents are updated in batches weekly or monthly.
Documents are divided into two categories: structured and unstructured. Structured documents include fields such as lease principal, annualized rate, lease start date, residual value rate, with units mostly being ten thousand yuan, percentage, year/month/day. Unstructured documents such as due diligence reports are mostly long texts, containing detailed content like leased asset evaluation, repayment plans, and guarantor qualifications.
What constraints these characteristics impose on model access and configuration
The need for multi-field matching with structured data requires models to support precise recall and semantic understanding of specified fields. Corresponding index rules for these fields must be configured.
Differences in update rhythms across data sources require configuration for switching between incremental and full synchronization, to avoid duplicate synchronization or missed updated content.
Variations in the length of unstructured documents require configuring segment lengths to fit the semantic integrity of individual documents, while controlling context window usage.
Professional terms such as sale-leaseback, leverage ratio, margin ratio require embedding models to have semantic recognition capabilities in the financial leasing field. Targeted adjustments to embedding configuration are needed.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1000 characters | The core semantic units of lease contracts and due diligence reports are mostly 800-1000 characters. Too short segments will damage clause integrity, while too long segments will reduce recall accuracy |
recall_top_k | 10–15 entries | A single financial leasing project is associated with many valid data entries. An overly high value will distract the model, while an overly low value will miss key guarantor or rent clause information |
maxContext | 8000–12000 characters | The length of a single due diligence report is mostly 5000-10000 characters. Sufficient context space must be reserved to accommodate recall results and query instructions |
json_schema_validation | Enabled | Investment research reports need to return structured rent calculation and leased asset evaluation results. Enabling this setting forces the model to output content that conforms to a preset format |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Parsing long individual documents takes a long time. 300 seconds covers the parsing process for most due diligence reports, preventing task timeouts and failures |
embedding_model | High-dimensional vector model | To meet the semantic matching needs of financial leasing professional terms, high-dimensional models can improve the recall accuracy of professional vocabulary |
The parameter values provided on this page are common recommended starting points for determining configurations. 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 settings.
Three common configuration errors
- Symptom: Knowledge base search returns results that cannot be parsed into standard JSON format. Cause: The
json_schema_validationconfiguration is not enabled, or the configured model does not support structured output. - Symptom: The tool call process cannot automatically select a matching tool. Cause: A model type that supports tool calling is not configured for
tool_call_model, or clear tool description fields are not configured. - Symptom: Mixed use of different types of embedding models leads to inconsistent recall results. Cause: The vector dimensions and recall logic of the embedding models are not unified, leading to incompatible vector spaces across different models.
How to confirm the configuration is complete
- A single financial leasing due diligence report is uploaded. The parsed segment length is checked against the configured
chunk_size, and the configuration is adjusted until the segments have complete semantics. - An investment research query involving rent calculation and leased asset evaluation is submitted. The returned results are checked for conformity to the preset JSON format, to verify that
json_schema_validationis active. - The tool call flow is tested. The model is confirmed to automatically select the corresponding tool based on the query content, to verify the rationality of the
tool_call_modelconfiguration. - The recall results of different embedding models are compared. Vector dimensions and recall logic are confirmed to be unified, to avoid result deviations.
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.