What This Category of Data Looks Like
Data sources for this category include enterprise internal control systems, ERP, financial sharing systems, audit management systems, and procurement contract project systems. The update frequency is annual. Daily documents cover internal control self-assessment reports, audit notices, audit working papers, issue lists, rectification ledgers, and rectification closure certificates. Documents contain clear business-related fields such as issue number, rectification responsible person, and closure status. Single document content ranges from hundreds to tens of thousands of characters. Some structured fields require cross-system association and matching.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Heterogeneous data from multiple systems requires retrieval to adapt to different document formats and structured fields, avoiding reliance on only a single retrieval logic. The annual update frequency means the knowledge base does not need frequent incremental synchronization, and full updates can be performed on an annual cycle to reduce operation and maintenance costs. Documents have clear issue-rectification associated fields, so retrieval must support associative recall of relevant rectification records and supporting materials. Document content spans a wide range and focuses on professional terminology, so precise filtering of low-relevance results is required, while avoiding missing key audit working papers and rectification vouchers.
How to Set the Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Audit-related documents have high requirements for result accuracy; this range can filter low-relevance non-audit content |
limit | Top 10–15 results | Audit issues need to be associated with multiple working papers and rectification records. Excessive results will exceed the context capacity limit and affect subsequent processing |
searchMode | Semantic + keyword hybrid retrieval | Audit documents contain both professional audit terminology and structured number fields. Hybrid retrieval can match both semantic and precise field requirements |
usingReRank | Enabled | Results recalled from multi-source heterogeneous data require secondary ranking to raise the priority of high-relevance audit working papers and rectification ledgers |
datasetSearchUsingExtensionQuery | Enabled | The wording of audit issues may not fully match document titles. Extended queries can cover more relevant historical audit documents |
collectionIds | Specify corresponding collections by business domain | Isolate audit data from different subordinate enterprises or departments to avoid cross-domain data interfering with retrieval results |
Reference only; verify against the target FastGPT release and deployment documentation.
Three Common Misconfigurations
- When calling the API, images uploaded to the knowledge base cannot be displayed. The cause is that the external access path for knowledge base attachments is not configured, and only local file paths are retained, making loading impossible in third-party systems.
- A "Cannot convert undefined or null to object" error occurs during workflow runtime. The cause is that no valid
collectionIdsparameter is specified in the retrieval configuration, causing an empty collection parameter to be passed during the call, triggering a type conversion error. - The retrieval context upper limit is set to 1500 characters, but document chunks exceeding the length are still recalled. The cause is that no filtering is performed on the length of individual document chunks, only the total context length is limited, or the chunk truncation logic is not enabled in the configuration.
How to Confirm the Configuration Is Complete
- Authorized personnel log in to the FastGPT backend, enter the retrieval configuration page of the corresponding knowledge base, and check that the values of parameters such as
similarity,limit, andsearchModematch the preset configuration. - A typical audit question is entered to trigger a retrieval test, and the recalled results are checked to confirm they include target document types such as audit working papers and rectification ledgers, and that the number of results conforms to the
limitsetting. - A document chunk exceeding 1500 characters is uploaded, a retrieval test is performed, and it is confirmed that the returned document chunks do not exceed the preset individual chunk length limit if chunk filtering is configured.
- A workflow with associated retrieval is run, runtime logs are reviewed, parameter conversion errors are confirmed to be absent, and the retrieval result association logic is checked to meet business requirements.
All configuration values provided here are common starting points, and should be evaluated against the specific samples used in the actual deployment scenario.
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.