What the data for this category looks like
This category of data comes from enterprise internal control systems, ERP systems, financial sharing systems, audit management systems, and procurement contract project systems. The update frequency is annual. Daily documents include internal control self-evaluation reports, audit notices, audit working papers, issue lists, rectification ledgers, and rectification closing certificates. Most documents contain fields such as business sample numbers, audit issue descriptions, rectification measures, and supporting material association fields. The structure is primarily a mix of semi-structured and structured content.
Constraints imposed on the knowledge base retrieval and recall link
Annual updated data means the knowledge base does not need frequent refreshes. The volume of data imported in a single batch is large, so chunking granularity must be controlled to avoid retrieval redundancy. Documents contain structured fields and semi-structured text, so targeted combination of field-based retrieval and semantic recall is required. Association fields such as business samples and rectification supporting materials require retrieval to link corresponding vouchers and ledgers, avoiding isolated recall of issue items. The rigor of audit-related documents requires that recall results match the business scenario with no irrelevant content. Retrieval results must also be traceable back to original business system data.
How to set configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the rigor requirements of audit documents, filters low-relevance recall results |
limit | Top 6–10 results | Adapts to the context length of a single audit document, avoids exceeding model input limits |
searchMode | hybrid | Combines structured field retrieval and semantic recall, suitable for audit documents with mixed structures |
usingReRank | Enabled | Performs secondary ranking on initial recall results, improves matching accuracy for rectification ledgers and issue lists |
datasetSearchUsingExtensionQuery | Disabled | Prevents extended queries from producing recall results that deviate from audit business scenarios |
collectionIds | Specified by business domain | Only recalls knowledge base collections related to internal control and audit, narrows the scope of invalid retrieval |
Reference only; verify against the target FastGPT release and deployment documentation.
Three common mistakes
- Phenomenon: Knowledge base retrieval returns empty results or triggers error messages and 400 status codes. Cause:
collectionIdsis not specified, the retrieval scope covers irrelevant knowledge base collections, or thesimilaritythreshold is set too high, resulting in no matching results. - Phenomenon: The parsing status cannot be obtained after calling the API to create a knowledge base. Cause: The parsing status synchronization logic is not configured correctly, or the corresponding callback interface is not monitored.
- Phenomenon: The semantic similarity value of retrieval results is abnormally high. Cause: A similarity calculation method outside the 0-1 range is misused, and the output format of the vector model is not aligned.
How to verify correct configuration
- Manually input an issue description from an audit working paper, and check whether the retrieval results include the rectification ledger and supporting materials of the same batch.
- Call the API to obtain the knowledge base parsing status, and confirm that the returned fields include three statuses: parsing in progress, ready, and parsing failed.
- Adjust the
similaritythreshold, observe changes in the matching degree of retrieval results, and confirm that the threshold meets business rigor requirements. - Test the full workflow, and confirm that the AI conversation step can be entered normally after the knowledge base search link.
The values provided are common starting points, and should be evaluated 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.