What Data for This Category Looks Like
Data for this category mainly 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 documents include audit notices, inspection work drafts, suspicious point lists, interview records, handling decision letters, and rectification reports. Most are structured forms and semi-structured documents. Fields cover violation types, handling process nodes, involved matters, and other items. There is no unified fixed unit. The update frequency is monthly.
Constraints Imposed by These Characteristics on the "Knowledge Base Retrieval and Recall" Link
Monthly updated data must adapt to batch synchronization processes to avoid missing changes to rules and settlement data during incremental synchronization. Mixed multi-type document structures require retrieval models to adapt to weight matching for different fields. For example, keywords in suspicious point lists and compliance descriptions in handling decision letters require separate recall logic. Medical insurance settlement data and drug and consumable traceability codes have many associated fields, so cross-data source associated recall must be supported during retrieval. At the same time, the timeliness requirement of audit documents requires that recall results match the latest updated supervision rules, and expired clauses must not be used.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Semantic similarity between audit suspicious point keywords and rule text. A value too low will recall irrelevant data, while a value too high will miss potential violation clues |
limit | Top 8 entries | Adapts to the volume of documents retrieved for a single audit task, balances retrieval efficiency and result coverage |
searchMode | keyword + semantic | Balances fixed-format keyword matching for audit notices and semantic description recall for the supervision rule base |
usingReRank | Enabled | Reranks initial recall results by violation relevance, improves the display priority of key suspicious points |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval terms to cover associated fields of medical insurance settlement data, improves the accuracy of cross-data source recall |
collectionIds | Specify corresponding collection IDs for the audit rule base, settlement data, and suspicious point list | Divides collections by business data classification, enables targeted retrieval for audit scenarios, and avoids interference from irrelevant data |
Three Common Misconfigurations
- Knowledge base export fails with a server error. The cause is that incremental synchronization tasks are not configured according to the monthly update rhythm, leading to export files containing oversized settlement data blocks that have not completed synchronization.
- Retrieval results fail to fully cover supervision rule content in the knowledge base. The cause is that
datasetSearchUsingExtensionQueryis not enabled, no extended matching is performed for retrieval terms, and only a small number of literal matching entries are recalled. - Retrieval tasks time out. The cause is that the number of recalled entries for the
limitparameter is not restricted, and too many cross-data source document data is loaded in a single retrieval, exceeding system processing limits.
How to Confirm Proper Configuration
- Run a single audit scenario retrieval test, verify that recall results include content from the corresponding rule base, settlement data, and suspicious point list.
- View the knowledge base synchronization log, confirm that the monthly incremental update task completes normally with no failed entries.
- Simulate exporting the full knowledge base file, check that the download process has no server error prompts.
- Adjust the
similarityparameter range, compare changes in the coverage of violation clues in recall results, confirm that the parameter takes effect.
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.