What data for this category looks like
This category of data comes from four sources: content review backend, user reporting channels, public regulations and policies, and annotation sample libraries. Daily maintained documents include review rule entries, disposal operation guides, appeal review records, and quality inspection review reports. Data updates are triggered by policy changes or synchronized daily. Document fields include rule numbers, effective time periods, violation judgment criteria, sample annotation tags, and more. Basic units are rule entries, record counts, and character counts.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source and frequently updated data requires the retrieval pipeline to support near-real-time synchronization, ensuring returned rules and records are the currently effective versions. Diverse document structures require adaptation to text blocks of varying lengths, to avoid truncating key rule effective conditions or judgment logic. Effective time periods in fields require adding time range filtering during retrieval to exclude expired rules. Annotation data in sample libraries requires precise matching of semantic associations, to avoid mistakenly recalling irrelevant old versions of samples. At the same time, synonym and industry variant expressions of rules must be covered to improve retrieval coverage during difficult sample judgment.
How to set configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to the precise matching requirements of review rules, filters low-relevance redundant results, and covers semantically similar rule expressions |
limit | Top 10 entries | Controls the number of returned results, adapts to the scenario of quick reference during on-call maintenance, and avoids information overload |
searchMode | semantic | Targets the logical associations of review rules, matches the context associations of rule explanations and difficult samples, and is not limited to keyword matching |
usingReRank | Enabled | Performs re-ranking on initial recall results, prioritizes returning high-confidence review rules and disposal guides |
datasetSearchUsingExtensionQuery | Enabled | Expands relevant keywords based on user questions, covering synonym and industry variant expressions of review rules |
collectionIds | Configured separately for the review rule library and appeal record library | Accurately locates the knowledge base collection exclusive to this category, avoiding mixing in irrelevant data from other categories |
Three common mistakes
- Phenomenon: Knowledge blocks containing specified keywords in the knowledge base are not recalled during semantic retrieval. Cause:
datasetSearchUsingExtensionQueryis not enabled, or thesimilaritythreshold is set too high, filtering out rule entries that have semantic matching but weak keyword association. - Phenomenon: When using a strict question-and-answer template, image address content in the knowledge base is retrieved, and the system prompts that no answer is found. Cause: The strict question-and-answer template only supports text-based knowledge base content, and no adaptive parsing rule for image content is configured.
- Phenomenon: API calls return a knowledge base retrieval failure, prompting that the knowledge base ID is invalid. Cause: The
collectionIdsparameter is configured incorrectly, the collection ID of the knowledge base exclusive to this category is not used, or the global variable is not correctly bound to the corresponding knowledge base.
How to confirm the configuration is complete
- Initiate a test query containing review rule keywords, and check whether the returned results include currently effective rule entries.
- Adjust the
similaritythreshold, observe changes in the relevance of recall results, and confirm that the threshold adapts to business requirements. - Call the API interface with the specified
collectionIds, and check whether the returned results only include knowledge base content of this category. - Switch between the strict question-and-answer template and other templates, and verify whether the retrieval results of image address content meet expectations.
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 2026-10-19.