Reference only; verify against the target FastGPT release and deployment documentation.
What This Category of Data Looks Like
Data for this category comes from government portals, document management systems, government information disclosure directories, policy document repositories, and press conference materials. Daily documents include policy documents, policy interpretation drafts, public disclosure directories, information release approval forms, press releases, and website column update records. The data update frequency is daily. Most document structures are structured official document paragraphs and entries with release document numbers and public attribute tags. Fields include document title, issuing unit, effective date, public scope, original content, and more. Some documents include approval process records. There are no unified standard fields, and fields vary across documents from different sources.
Constraints Imposed on Knowledge Base Retrieval and Recall
Daily incremental updates require the retrieval system to support incremental synchronization and incremental indexing, avoiding resource consumption from full reindexing. Diverse document formats and lack of unified standard fields require the retrieval link to support multi-format text parsing and unstructured content recall. Metadata fields related to public attributes and confidentiality reviews require retrieval results to be associated with public scope tags, filtering content that does not meet public disclosure requirements. The strong correlation between policy interpretations and original documents requires recall results to cover related documents, avoiding misinterpretation of policy stances. Daily update timeliness requires retrieval latency to be controlled within a reasonable range, ensuring timely access to public information.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.6–0.75 | Matches the semantic relevance between policy documents and user queries, filters low-relevance results, and ensures the accuracy of public information |
limit | Top 8 | The number of documents recalled per single request adapts to the volume of government public disclosure content, preventing excessive results from causing context overflow |
searchMode | vector | Adapts to the semantic retrieval needs of policy documents, prioritizes matching content relevance, uses semantic matching logic, and does not use exact keyword matching |
usingReRank | Enabled | Performs secondary ranking on recall results, corrects deviations in semantic retrieval, and improves the correlation matching between policy interpretations and original documents |
datasetSearchUsingExtensionQuery | Enabled | Generates extended retrieval terms based on user queries, covering synonyms and abbreviations of policy documents, and adapts to the diversity of government terminology |
collectionIds | Specified by public document collections | Limits the retrieval scope to only compliant documents such as proactively disclosed policy files and interpretation drafts, avoiding retrieval of non-public content |
Three Common Misconfigurations
- Phenomenon: Semantic retrieval scores show values over 4000, exceeding the conventional 0-1 range. Cause: No normalization rule for semantic scores is configured, resulting in uncalibrated scores from original embedding vector calculations, which does not comply with retrieval result display specifications for government scenarios.
- Phenomenon: No search results are returned during debugging, or only a small number of irrelevant content is returned. Cause:
collectionIdsdoes not specify the correct proactively disclosed document collection, orsearchModeis configured to a mode incompatible with government document formats, resulting in incorrect retrieval scope coverage. - Phenomenon: Source information from knowledge base searches is forcibly displayed in responses. Cause: The display switch for retrieval results is not turned off, or the default retrieval source output is retained in the configuration, which does not comply with the requirement for concise responses in government public disclosure scenarios.
How to Verify Proper Configuration
- Perform a policy query, check whether the document sources of the recall results are the specified proactively disclosed collection, and confirm that the
collectionIdsconfiguration takes effect. - Check the display range of retrieval scores, confirm that the scores are within a reasonable range, and verify that the semantic score normalization configuration is correct.
- Trigger a retrieval, check whether the response contains redundant search source information, and confirm that the retrieval display switch is configured according to scenario requirements.
- Simulate daily incremental update document upload, verify whether the retrieval system can recall newly uploaded public documents in real time, and confirm that the incremental synchronization configuration takes effect.
The values given 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.