What data in this category looks like
Data in this category originates from permission-controlled institutional documents and business data documents within central state-owned enterprises and government departments. This includes official documents with red headers, internal process specifications, permission-bound business reports, and similar materials. Data updates occur irregularly, triggered by system revisions and business data updates. A single document typically includes structures such as a permission scope field, release document number, applicable position or department, and main body content. The main data units are characters, file size in MB, and permission codes.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The permission scope field requires the retrieval link to bind user permission identifiers, filtering recall results that exceed user permissions to avoid sensitive information leakage. The irregular update feature requires the retrieval system to support incremental synchronization configuration, triggering index updates only when data changes to reduce redundant calculations. Documents contain fixed structure fields, which can be used to accurately filter the recall scope. Long text content must also adapt to segmentation rules, preventing single segments from being too long and exceeding context limits. The file sizes of permission-controlled files vary widely, so processing must adapt to the character length of different files to ensure retrieval efficiency.
Reference only; verify against the target FastGPT release and deployment documentation.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to the semantic similarity of internal system texts, balances retrieval precision and coverage, and avoids missing compliant content |
limit | Top 10 entries | Internal system documents usually have standardized structures. 10 recall results can cover the context needs of common Q&A, avoiding excessive redundant information |
searchMode | vector | Internal systems mainly rely on text semantic matching, and vector retrieval can accurately match the relevance between system clauses and user questions |
usingReRank | Enabled | Reranks recall results, optimizes the sorting logic after long document segmentation, and ensures that core system clauses are displayed first |
datasetSearchUsingExtensionQuery | Disabled | Internal system expressions are rigorous. Extended queries may introduce irrelevant content, maintaining the precision of original queries |
collectionIds | Configured by permission groups | Binds the collection ID of the user's department or position, only retrieves knowledge base collections within the corresponding permission scope, achieving data isolation |
Three common misconfigurations
- After adding data to a collection, no usable index is generated, and empty results are returned when attempting to query. Index construction is not triggered by executing a training order; only data is added without completing vectorization and index generation.
- Retrieval results include document content that exceeds user permissions. No permission filtering logic is configured, and
collectionIdsor user permission identifiers are not bound to retrieval parameters. - The number of retrieved results does not match expectations, either too many or too few. The
similaritythreshold is set unreasonably, or thelimitvalue does not align with the recall needs of the current scenario.
How to confirm the configuration is valid
- Upload a test document within the permission scope, initiate retrieval using an account with corresponding permissions, and verify that the document content appears in the results. Confirm that no relevant results are returned when an account without the required permissions retrieves.
- Trigger a data update, check the index update log, and confirm that only newly added or modified documents are processed, with no full index reconstruction triggered.
- Initiate a test query, review the number and sorting of returned results, and confirm they align with the configuration logic for
limitandusingReRank. - Check the retrieval request parameters, confirming that
collectionIdsis bound to the permission group collection ID of the current user to ensure data isolation takes effect.
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.