What this category of data looks like
This category of data mainly comes from central enterprise internal control systems, ERP, financial sharing systems, audit management systems, and procurement contract project systems. Daily produced documents include internal control self-assessment reports, audit notices, audit working papers, issue lists, rectification ledgers, and rectification closing certificates. The data update frequency is annual. Most documents contain fields such as business sample number, voucher identifier, issue description, rectification requirements, and supporting material number. Some documents have both structured tables and unstructured attachments, and the size of individual texts varies widely.
What constraints do these characteristics impose on the citation source and traceability link
Data sources are scattered and connected to multiple business systems. It is necessary to accurately match field identifiers of different systems to avoid traceability mismatches. The low-frequency annual update attribute requires the traceability mechanism to retain historical version association capabilities, to adapt to audit traceability requirements. Diverse document types including structured and unstructured content require support for cross-format source location. The strong association between rectification ledgers and closing certificates requires the traceability link to bind issues and corresponding supporting materials, ensuring the auditability of the rectification process. For the time-consuming step of extracting business samples, it is necessary to quickly locate associated records of original vouchers and audit working papers, to reduce manual verification costs.
How to set up the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Adapts to the text length of single documents such as audit working papers and rectification ledgers, avoiding truncation of key supporting information |
responseData | Enabled | Returns original cited data for audit traceability, to meet regulatory verification requirements |
detail | Full | Displays the complete path and field information of the citation source, including system source, document number, and upload time |
| Single recall limit | Top 5 entries | Controls the number of citations per round of responses to avoid information overload, aligning with the conciseness requirements of audit reports |
| Context recall range | Data from the past 12 months | Matches the annual update frequency of this category, avoiding the introduction of expired invalid data |
| Citation display format | Structured list | Clearly labels document types and unique identifiers for easy manual verification |
Three common mistakes
- Generated citation formats are messy, with meaningless symbols or truncated content. Cause: The
quoteMaxTokenparameter is not configured correctly, causing cited text to be abnormally truncated or spliced after exceeding the limited length. - Displayed citation content has garbled characters. Cause: The encoding formats of multiple system data sources are not adapted, or the character set processing logic is not unified when recalling citation data.
- Knowledge base cited content is not displayed in the main text. Cause: The
responseDataparameter is not enabled, causing the system to only return response text without attached citation source data.
How to confirm the configuration is correct
Reference only; verify against the target FastGPT release and deployment documentation.
- View the specified area of the response text to confirm that there is a structured citation list labeling document types and unique identifiers.
- Test inputting a document fragment exceeding 800 tokens, check whether the cited content is fully retained without abnormal truncation.
- Switch test data from different data sources, confirm that the recalled results only include valid data from the past 12 months, with no expired content introduced.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.