What this category of data looks like
Data for query result reports originates from monthly and quarterly business statistical reports submitted by internal departments of central state-owned enterprises and state-owned enterprises, plus data summary attachments included in internal system documents. Some reports are submitted in Word document format. The data update rhythm adjusts based on business needs. Regular updates occur monthly, while special business reports are generated on demand. Most documents consist of structured tables paired with explanatory text, with fixed fields including report number, generation time, affiliated business line, statistical indicator name, corresponding value and unit. Some long reports covering multiple business lines display detailed data across multiple pages.
What constraints these characteristics impose on the knowledge base retrieval and recall link
First, the sensitive nature of internal reports and systems requires retrieval to be bound to department permissions, preventing unauthorized access to data outside the user’s department. Second, the monthly update feature requires configuring incremental recall logic, reducing resource consumption from full retrieval and improving retrieval efficiency. The presence of structured fields and fixed units requires retrieval to support field-level precise matching, avoiding invalid results where values and units are mismatched. The structure of long paginated reports requires retaining complete table rows and columns during chunking, avoiding broken data semantics after splitting that reduces retrieval matching accuracy.
How to set up the configuration
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Match structured indicators and explanatory text of reports, filter low-correlation irrelevant content |
limit | Top 10–15 entries | Control the number of reports recalled in a single retrieval, avoid context window overflow |
searchMode | vector + keyword | Balance semantic vector matching and precise keyword matching, adapt to structured field retrieval needs |
usingReRank | Enabled | Perform secondary ranking on recalled multiple reports, ensure results related to core business indicators are displayed first |
datasetSearchUsingExtensionQuery | Disabled | Indicator names of internal reports are fixed, extended queries easily cause ambiguity and reduce retrieval precision |
collectionIds | Specify exclusive collection ID for business department | Bind department permissions to achieve isolated access to internal data |
Three common misconfigurations
- Phenomenon: Retrieval results include non-department report data, or permission error prompts appear. Cause: The
collectionIdsparameter is not configured correctly, and the exclusive department dataset is not bound, leading to cross-department data leakage or failure to comply with access compliance requirements. - Phenomenon: Detailed data of long reports is split, leading to broken semantics, and complete business indicators cannot be matched during retrieval. Cause: The complete row structure of the table is not retained during chunking, and page-by-page report details are split into independent chunks, losing key context-associated information.
- Phenomenon: Embedded charts and images in uploaded report documents in the knowledge base cannot be recalled normally, and no corresponding visual content or indicator descriptions appear in retrieval results. Cause: Extraction rules for non-text elements in documents are not configured, only plain text paragraphs are captured, and statistical indicator information bound to charts is lost.
How to confirm the configuration is correct
- Submit a test report exclusive to the department, initiate retrieval using the corresponding department account, and verify that recalled results only include reports from this department.
- Upload a test report containing a long table, check that chunking during retrieval retains the complete row and column structure of the table, with no split breaks.
- Adjust the value range of relevant parameters, initiate a query containing a specific indicator name, and verify that the matching degree of recalled results meets expectations.
- Enable the
usingReRankconfiguration, compare recall sorting results before and after enabling, and confirm that reports related to core business indicators are displayed first.
The values listed above are common starting points, and should be evaluated against the user’s own sample data.
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.