What this category of data looks like
Data governance teams, security consultants, and legal and compliance teams produce this category of data. Core documents include data catalogs, classification and grading rules, important data identification tables, and review minutes. Data sources include data asset inventory results, database metadata, business system descriptions, and compliance reference documents. The update frequency is quarterly or when a business system is changed on an ad hoc basis. Documents contain structured fields and unstructured text, with no unified fixed-length format, requiring adaptation for retrieval and matching across multiple content types.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Data sources include metadata and business system descriptions, leading to large differences in content formats. Teams must support both structured field matching and non-semantic text retrieval. The update frequency is quarterly or on-demand changes, so knowledge base retrieval must adapt to the low-frequency nature of incremental synchronization, avoiding resource consumption caused by frequent full updates. Documents include semi-structured content such as review minutes, so semantic similarity judgment requires more precise calibration. In addition, due to data association compliance requirements, retrieval results must strictly match preset classification and grading rules, and no irrelevant content that interferes with compliance judgment can appear.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Classification and grading rules are rigorous compliance text, requiring filtering of irrelevant retrieval results with low matching degrees |
limit | Top 8 entries | Single classification and grading document has relatively long content, and excessive recall will exceed the context window limit |
searchMode | vector+keyword hybrid mode | Covers retrieval needs for structured metadata fields and unstructured rule text |
usingReRank | Enabled | Semantic similarity judgment for semi-structured content such as review minutes requires secondary calibration to improve ranking accuracy |
collectionIds | Bind exclusive knowledge base collection ID list | Isolate classification and grading data from other business data to avoid retrieval interference |
datasetSearchUsingExtensionQuery | Enabled | Convert user questions into compliance retrieval sentences matching data catalogs and identification tables, improving recall accuracy |
Three Common Misconfigurations
- Using raw request data returned by the HTTP interface as a knowledge base reference results in empty retrieval results. The cause is that the raw data fails to align with the preset field structure of the data catalog and classification and grading rules, so it cannot be correctly parsed by the embedding model.
- No results are returned for the first retrieval of classification and grading-related questions, but normal results are returned for the second retrieval. The cause is a delay in knowledge base incremental synchronization. Since the update frequency of this category of data is quarterly or on-demand changes, the synchronization trigger mechanism must be adjusted to match the update rhythm.
- Multi-knowledge base retrieval does not execute in the preset order, and the second knowledge base is not triggered after no results are retrieved from the first one. The cause is that the incoming order of
collectionIdsdoes not follow business priority, or the sequential retrieval logic for multiple knowledge bases is not set up.
How to Verify Configuration Validity
- Upload a single classification and grading rule document, run a retrieval to match core compliance clauses in the document, and confirm that retrieval results include the target content.
- Adjust the
similarityparameter value, observe changes in retrieval result matching degrees, and confirm that the parameter configuration takes effect. - Pass in an ID from a non-preset knowledge base, confirm that retrieval returns empty results, and verify that the
collectionIdsisolation configuration takes effect. - After enabling
usingReRank, compare the ranking of retrieval results before and after reranking, and confirm that the reranking function operates normally.
The values provided are common starting points and should be measured against relevant samples in use.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.