What the data for this category looks like
Data for this category comes from the medical insurance intelligent monitoring system, fund supervision rule base, medical insurance settlement data, drug and consumable traceability codes, and complaint clues. Daily processed documents include audit notices, inspection work papers, suspicious point lists, interview records, decision letters, and rectification reports. Most documents include fields such as designated medical institution name, expense details, and basis for violation determination. Units involved include yuan, number of visits, and drug batch numbers. Data updates occur once per month. Overall data accumulates gradually with each audit batch.
Constraints Imposed on Knowledge Base Retrieval and Recall
Data sources are scattered. Structured and unstructured data from multiple channels must be integrated. Retrieval must cover multi-dimensional associated fields. The monthly update requirement means the knowledge base sync cycle must match the data update rhythm. Delayed data must not interfere with violation determination. Document types are varied. Both standardized rule base text and unstructured interview records exist. Retrieval weights must be adjusted for different document types. Audit workflows require quick access to associated files. Retrieved results must accurately match violation clues in suspicious point lists. Redundant results must not disrupt manual review.
Configuration Recommendations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the relevance between medical insurance violation determination rules and files. Filters low-correlation non-violation documents, and aligns with the score range of FastGPT v4.17.0 |
limit | Top 8–12 results | Controls the number of returned results per retrieval batch. Adapts to manual review efficiency in the audit workflow, and prevents context overflow |
searchMode | keyword + vector hybrid | Balances keyword matching for standardized rule base text and semantic association for unstructured files, covering all document types |
usingReRank | Enabled | Performs secondary ranking on initial retrieved results. Prioritizes results that match violation determination keywords in suspicious point lists |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms based on audit suspicious points. Covers vague descriptions in complaint clues |
collectionIds | Exclusive knowledge base collection ID for this audit scenario | Isolates documents from other categories, preventing irrelevant data from interfering with retrieval accuracy |
Three Common Configuration Errors
- The number of results returned after retrieval does not match the configured
limitvalue. This occurs whencollectionIdsis not properly configured, causing the retrieval scope to include knowledge base data from non-audit scenarios. - The "select knowledge base" parameter in global variables does not take effect after being passed. This happens when an incorrect variable format is used, and the required parameter structure specified by the system is not followed.
- The interface prompts a format error when uploading knowledge base files. This is because the uploaded documents do not match the supported format types, or the files contain unparseable special characters.
How to Verify Successful Configuration
- Log in to the FastGPT v4.17.0 management backend, navigate to the retrieval configuration page for the target knowledge base, and confirm that all configuration item values match the preset plan.
- Enter a typical audit suspicious point, run a retrieval, and confirm that the number of returned results matches the range specified by the
limitconfiguration. - Manually trigger a knowledge base sync task, and confirm that the sync progress aligns with the monthly data update cycle.
- Call the retrieval interface, and check whether the associated fields in the returned results match the input audit suspicious points.
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.