What this category’s data looks like
The core data of this category comes from the national basic catalog of government service items, administrative licensing item list, power responsibility list, and local government service item databases. Daily maintained documents include the basic catalog of government service items, implementation lists, service guides, standard material directories, flowcharts, and frequently asked questions. Data updates occur monthly. Document structures include standardized fields such as item code, item name, acceptance conditions, application material list, and handling process. Some documents include visual flowcharts and Q&A text.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The monthly update frequency requires the knowledge base to support regular full synchronization. This avoids catalog version inconsistencies caused by incremental synchronization. Documents include structured fields and visual flowcharts. Retrieval and recall must cover both text content and associated matching of unstructured resources. The standardized field system requires the retrieval system to support precise field filtering. It must also ensure recall accuracy for core fields such as acceptance conditions and application materials. Maintenance work requires comparing upper and lower level catalogs. Recall results must be traceable to their data source to ensure unified service standards online and offline.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets precision requirements for matching government service item fields, and avoids low-relevance results |
limit | Top 10 entries | Covers the full volume of items updated monthly, ensuring retrieval results include all catalogs to be compared |
searchMode | Semantic + structured hybrid retrieval | Adapts to the characteristics that item documents contain both text descriptions and structured fields |
usingReRank | Enabled | Improves ranking accuracy for high-frequency item retrieval, and corrects local deviations in basic similarity scores |
datasetSearchUsingExtensionQuery | Enabled | Supplements synonym expansion for item names, covering alternative terms used in local government services |
collectionIds | Exclusive collection ID corresponding to the government service item database | Isolates other business data, ensuring the retrieval scope only covers documents in this category |
Three Common Mistakes
- Phenomenon: Entries with a score of 1.2717 rank lower than entries with a score of 0.5208 in retrieval results. Cause: The
usingReRankreranking function is not enabled. Sorting only relies on basic similarity scores without correcting local semantic matching deviations. - Phenomenon: The knowledge base remains in the indexing state for a long time after upload. Cause: Reasonable batch indexing thresholds are not configured for the large-volume government service item database updated monthly, leading to backlogged indexing tasks.
- Phenomenon: Flowchart images cannot be displayed in retrieval results. Cause: Associated recall configuration for unstructured resources is not enabled. Only text content is retrieved without associating bound image files.
How to confirm the configuration is correct
- Run a single item name retrieval. Check if returned results are sorted by semantic relevance to verify the
usingReRankreranking function is active. - Upload a single test service guide document. Confirm indexing progress completes within 10 minutes to troubleshoot indexing backlog issues.
- Retrieve items that include acceptance condition fields. Verify that field content can be matched precisely to confirm the
searchModehybrid retrieval mode is active. - Retrieve items that include flowcharts. Check if returned results include associated image resources to verify unstructured resource recall configuration.
The values listed above are common starting points, and should be evaluated against the user’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.