What the data for this category looks like
Data for joint-stock bank intelligent due diligence reports comes primarily from internal credit approval archives, due diligence copies from peer institutions, public financial report disclosures, and regulatory submission documents. Updates trigger synchronously with credit project initiation, quarterly financial report disclosures, and regulatory document releases. There is no fixed update cycle. Each single report includes structured fields such as subject qualifications, credit limits, guarantee clauses, financial indicators, and risk assessments. It also includes thousands of words of due diligence explanations and risk warning long texts. All fields follow internal bank risk management document specifications uniformly.
What constraints these characteristics impose on vector models and indexing
Many structured fields have fixed units. Vector models must distinguish encoding logic for numerical and text data to avoid similarity calculation deviations caused by unit ambiguity. Updates follow no fixed cycle. Indexes must support incremental updates without full reconstruction to reduce computing resource usage. Each report contains thousands of words of long text. Segmented indexing must adapt to long-text sliding window strategies to avoid truncating core risk assessment content. Minor format differences exist across data from multiple sources. Indexes must support field mapping rules for different documents to prevent structured fields from being lost.
How to set configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the long-text structure of due diligence reports, preserves complete risk assessment logic, and prevents key information from being truncated |
chunk_overlap | 100–150 characters | Covers core content at segment boundaries, prevents loss of contextual association after long text segmentation |
embedding_model | Models with maximum token count ≥ 8192 | Adapts to the thousands of words of long text in due diligence reports, avoids truncation of core risk information during encoding |
retrieval_top_k | Top 8–12 results | Covers multi-dimensional risk points required for due diligence, balances recall comprehensiveness and result accuracy |
similarity_threshold | 0.72–0.80 | Meets the rigor requirements of risk management scenarios, balances similarity matching accuracy for structured fields and long texts |
enable_incremental_index | Enabled | Adapts to the non-fixed update cycle of due diligence reports, reduces computing resource consumption from full index reconstruction |
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 individual analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: After deploying a privatized reranking model in version 4.9, calls return empty results. Cause: Correct privatized model interface address and authentication parameters were not filled in the
RETRIEVAL_RERANK_MODELconfiguration item. - Phenomenon: Retrieval results fail to cover all knowledge base content, with a large number of missed recalls. Cause: The
similarity_thresholdwas set too high, filtering out some eligible due diligence report segments. - Phenomenon: After switching indexing models, the relevance of retrieval results does not change as expected. Cause: The
embedding_modelconfiguration item was not updated synchronously, and documents were still encoded using the old vector model.
How to confirm correct configuration
- Upload a test due diligence report, check the segment preview in the console, and confirm that the segment length matches the
chunk_sizeandchunk_overlapsettings. - Enter core risk keywords from the due diligence report to initiate retrieval, and verify that the number of recalled results matches the
retrieval_top_ksetting. - Add a new incremental due diligence report, wait for index updates to complete, then initiate retrieval, and confirm that the new document content can be recalled normally.
- Trigger a privatized reranking model call, and verify that the returned result ranking matches the ranking displayed in the console.
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.