What the data for this category looks like
Sanctions list screening data primarily comes from sanction lists published by national law enforcement agencies, industry-shared risk entity databases, and compliance risk records notified by regulators. Data update frequency varies by source: law enforcement agency lists are updated in real time or daily, while industry-shared databases are synchronized hourly. The document structure of a single record includes fields such as entity name, document type, document number, risk level, issuing authority, effective date, and expiration date. Risk levels are categorized as low, medium, or high. Dates use the YYYY-MM-DD format. Some records include additional fields for associated entities.
What constraints do these characteristics impose on multi-turn dialogue and prompt engineering
Decentralized data sources and frequent updates require the multi-turn dialogue module to support real-time pulling of the latest list data, and to avoid using expired cached data. A large number of fixed-structure fields require the prompt to clearly specify the range of fields to extract. During multi-turn dialogue, missing entity information must be confirmed step by step. The relatively long length of single records requires limiting the total character count of dialogue context, to prevent information loss from exceeding model window limits. Some records include aliases, so multi-turn dialogue must support an interactive step for verifying aliases.
How to Configure Parameters
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | First 8 turns of dialogue (approximately 4000 characters) | Sanctions list verification requires associating previously entered entity information. Excess old context will interfere with current field extraction and matching |
ragTopK | Top 6 matching results | Risk list entries are numerous in volume. Returning too many results will exceed model output limits, while returning too few may lead to missed high-risk entities |
promptTemplate | Fixed template: Please verify if the following entity is on the risk list. Extract fields including entity name, document type, document number, risk level, and issuing authority. Supplement verification if aliases exist | The list field structure is fixed. Clear prompts reduce the chance of the model missing required fields and standardize output formats |
contextRefreshInterval | Refresh risk list cache every 15 minutes | Law enforcement agency lists are updated at high frequency. Regular refreshes ensure the timeliness of verification data |
responseMaxTokens | 2000 characters | A single verification may return multiple matching records. Limiting output length prevents results from being truncated |
fileUploadMaxSize | 50 MB | Supports uploading complete list files with large numbers of risk entities to meet batch verification needs |
The parameter values provided on this page are common recommended starting points for configuration. 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 Misconfigurations
- Phenomenon: When the model outputs results in Markdown table format, the results are truncated and display
...[hide X char]. Cause: TheresponseMaxTokensparameter is not set, or its value is too small to accommodate complete matching result fields. - Phenomenon: After sending a verification request, the response returns
insufficient_quota 当前分组上游负载已饱和，请稍后再试. Cause: No risk list cache refresh mechanism is configured. Multi-turn dialogue frequently triggers real-time pull requests, exceeding system load limits. - Phenomenon: Verification is completed only based on the current single dialogue, without associating previously entered entity information. Cause:
maxContextis incorrectly set to 1, and key information such as document numbers and entity aliases confirmed in historical dialogue is not retained.
How to Verify Proper Configuration
- Upload a test risk list file, confirm that parsed fields fully cover required items such as entity name and document type, and check that segmentation rules do not miss key fields.
- Initiate a multi-turn verification dialogue, enter the entity name first, then supplement the document number. Confirm that the system can associate historical dialogue information to complete full verification, and that the context turn count setting meets requirements.
- Send multiple verification requests consecutively, confirm that the system regularly updates the risk list cache, and that the cache refresh cycle configuration is active.
- Trigger a verification request with multiple matching results, confirm that no truncation prompt appears in the output, and that the output length limit setting is reasonable.
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.