What Data for This Use Case Looks Like
The data for intelligent due diligence reports comes mainly from credit approval archives in internal credit management systems, financial statements of corporate clients, publicly disclosed documents from regulatory authorities, and internal risk control review records. Update rhythm follows the credit project cycle: real-time synchronization during the period from project initiation to archiving for individual projects, and quarterly batch updates for existing projects. The document structure is a combination of multiple fields, including basic customer information, financial indicators, risk control ratings, guarantee status, and regulatory compliance records. Fields include structured numerical values and unstructured review text, with units such as RMB yuan, rating levels, and other types.
Constraints Imposed by These Characteristics on Vector Models and Indexing
The mixed multi-source data structure of due diligence reports requires vector indexes to support joint vector mapping of structured numerical values and unstructured text, avoiding insufficient adaptation of single vector dimensions. The rhythm of real-time synchronization per project cycle and quarterly batch updates requires indexes to support flexible switching between incremental indexing and full reconstruction, adapting to efficiency requirements of different update scenarios. The large span in individual project document length requires segmentation strategies to adapt to differentiated processing of short fields and long texts, avoiding loss of key information caused by long text truncation. The existence of standardized field units and rating levels requires indexes to retain field metadata to support subsequent precise recall and result filtering.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the mixed length of long review texts and structured fields in due diligence reports, avoiding information overload or excessive fragmentation in a single segment |
vector_model | bge-large-zh-1.5 or a compatible model with an embedding dimension of 1024 | Adapts to the semantic capture accuracy of multi-source text, matches the text semantic features of due diligence reports, and complies with the model compatibility rules of FastGPT v4.8.7 |
index_incremental_update | Enabled | Adapts to real-time synchronization requirements per project cycle, reducing resource consumption from full index reconstruction |
retrieve_top_k | Top 8–12 entries | Covers associated information across multiple fields in due diligence reports, avoiding content missing caused by insufficient recall fragments |
similarity_threshold | 0.72–0.80 | Filters low-relevance unstructured text fragments, retaining content strongly related to due diligence objectives |
upload_file_max_size | 2000 MB | Adapts to the storage and indexing requirements of individual due diligence reports containing multiple attachments |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing.
Three Common Misconfigurations
- Phenomenon: The interface only displays a single vector model selection option, and it is not possible to configure independent vector models for different business fields. Cause: The field-level vector mapping switch is not enabled, and only a global single vector model is used to process all data, which cannot meet the differentiated semantic requirements of multiple fields.
- Phenomenon: Index tasks continuously show "running" in the background, with no progress updates for more than 2 hours. Cause: The
chunk_overlapparameter is not set, causing circular overlap after long text segmentation, or theupload_file_max_sizethreshold is set too low, causing repeated retries after large attachment upload failures. - Phenomenon: Locally trained vector data cannot be loaded normally on the server side, and recall results differ from local test results. Cause: The vector model used on the server side does not match the embedding dimension of the local model, or the model's word segmentation dictionary file is not synchronized.
How to Verify Proper Configuration
- Verify the embedding dimension parameter of the vector model, confirm that it matches the currently deployed model version, and adjust the model selection based on the text features of the due diligence report.
- Submit a small-volume due diligence report attachment, check whether the index task progresses to normal completion, and confirm the effective status of the incremental update configuration.
- Input test keywords related to the due diligence report, perform a recall test, and check whether the number of recall fragments and similarity results meet the business configuration requirements.
- Export the index configuration log, confirm that all enabled field-level vector mapping parameters have been correctly loaded, and there are no configuration conflicts.
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.