What this category of data looks like
Data mainly comes from three primary sources. These are identity verification materials submitted by claim applicants: resident ID cards, household registers, motor vehicle driving licenses, and similar documents. It also includes accident time-related certification documents: accident certificates, medical records. Timestamps for report filing and material submission are also included. Identity materials are collected once when a single claim process starts. Timing data is updated in real time as process nodes advance. Document structures mostly combine structured form fields and attached files. Fields fall into two categories. Identity information fields include name, ID number, and issuing authority. Timing fields include report time, material submission time, and accident time. Units used include date format (YYYY-MM-DD), millisecond-level timestamps, and document numbers stored as strings.
What constraints these characteristics impose on model access and configuration
Identity and timing data includes both structured fields and unstructured attachments. This creates a clear need for multi-modal model adaptation. Timing data uses mixed formats: timestamps and text-based time descriptions. Targeted time parsing rules must be configured. Timing data is updated in real time across process nodes for a single claim. Parameters that support low-latency calls must be configured. Identity information includes sensitive fields. Data desensitization trigger rules must be configured. Attached files are mostly document scans. Upload checks that comply with format and size limits must be configured. Material formats vary widely across different claim cases. General preprocessing parameters must be configured to avoid recognition failures from format incompatibility.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
multimodal_model_enable | Enabled | Identity materials are mostly document scans. A multi-modal model is required to complete text recognition and information extraction |
max_context_length | 8000–12000 characters | The setting must accommodate both identity field information and timing-related process text. This avoids truncating critical verification content |
timeout | 30 seconds | This aligns with the timing requirements of claim initial review. It avoids interrupting the initial review process due to call timeouts |
data_sensitive_mask | Enable ID number and name desensitization | This complies with compliance requirements for identity information processing. It prevents sensitive data leaks |
UPLOAD_FILE_MAX_SIZE | 10 MB | This covers the file sizes of most common document scans. It avoids blocks from overly large files |
reranker_top_k | Top 3 entries | Timing-related associated data volume is limited. A small number of retrievals is sufficient to complete accurate verification |
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 specific analysis. It is recommended to test on your own samples before finalizing values.
Three common configuration errors
- Phenomenon: When calling the
bge-rerankermodel, GPU video memory grows rapidly to 6-7G with each call, and multiple cards are occupied at the same time. Cause: Thereranker_batch_sizeparameter is not set, and the default batch processing logic does not limit concurrent request volume. - Phenomenon: Model access configuration cannot initiate normal calls after saving, returning an authentication failure error. Cause: Required fields
api_keyandendpointare not filled correctly, and the format does not meet platform requirements. - Phenomenon: Key fields are missing from identity material recognition results, such as the ID number not being extracted. Cause: The multi-modal model configuration is not enabled, and only a plain text model is used to process document scans.
How to confirm the configuration is complete
- Upload a document scan, trigger a model call, and verify the completeness of the returned identity fields. Adjust multi-modal related configurations until the recognition results meet requirements.
- Input timing data that includes both timestamps and text descriptions, and verify the model's parsing results. Adjust the context length configuration to ensure complete content is read.
- Upload document attachments of different formats and sizes, and verify the upload check results. Adjust the file size limit configuration to cover common scenarios.
- View call logs, verify the transmission of authentication parameters, and adjust access configuration items to ensure normal authentication.
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.