What this category of data includes
Data for this category originates from the medical insurance intelligent monitoring system, fund supervision rule base, medical insurance settlement data, drug and consumable traceability codes, and complaint clues. The update frequency is once per month. Daily processed documents include audit notices, inspection work papers, suspicion point lists, interview records, handling decision letters, and rectification reports. Each document contains fields such as business subject information, suspected violation descriptions, handling basis, and rectification requirements. There is no unified standard unit, so corresponding business fields must be matched based on document type.
Constraints from these characteristics for vector models and indexes
Bulk data updated monthly requires indexes to support switching between incremental synchronization and full reconstruction. This avoids reindexing full datasets and reduces processing time. Documents vary widely in structure across types. Some documents contain long-text violation clues, so splitting logic must be adapted for different fields. This prevents key information from being lost due to long text truncation. Massive unstructured medical and supervision terminology text from medical insurance settlement data requires vector models to adapt to Chinese professional scenarios. Indexes must also support filtered recall based on document type and time range, to meet the precise positioning needs of audit scenarios.
Configuration settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Covers the complete description of a single suspected violation point, avoids splitting breaks that lose key clues |
| Overlap Length | 100–200 characters | Retains contextual connections across split chunks, ensures complete semantics of suspected violations |
| Similarity Threshold | 0.75–0.85 | Filters low-relevance non-violation text, retains highly matched supervision clues |
| Single Search Limit | Top 20 entries | Covers multi-dimensional suspicion sources required for audits, avoids excessive recall results affecting analysis |
maxToken | 4096 | Adapts to the input length limit of mainstream Chinese medical supervision scenario vector models such as bge-large-zh-1.5 |
defaultToken | Calibrated based on actual testing | Controls the text processing length of a single batch index, adapts to the processing capacity of the system queue |
Three common configuration mistakes
- Phenomenon: Vector recall results do not match locally split documents, or a large number of irrelevant results appear. Cause: Unified vector model and splitting parameters are not used. Embedding dimensions or text processing logic differ across models, leading to inconsistent vector spaces between indexes and queries.
- Phenomenon: A single file remains in the indexing state for a long time with no progress updates. Cause: Reasonable single-file processing threshold is not set. Oversized files after splitting exceed the processing limit of the indexing queue, or the file contains unparseable binary content.
- Phenomenon: Recall results are not filtered according to the preset knowledge base priority, and irrelevant documents across knowledge bases are recalled first. Cause: Index knowledge base tag filtering rules are not configured, or priority parameters are not specified during retrieval, leading to index recall not sorted by business scenario priority.
How to verify correct configuration
- Upload a standard audit notice document, check the number of split chunks after indexing is complete, confirm the split length falls within the preset range.
- Call the retrieval interface, input a suspected violation description, check if the similarity scores of recall results fall within the preset threshold range.
- Upload an oversized file, verify that indexing completes within a reasonable time frame with no stuck state.
- After configuring multiple knowledge base tags, specify a specific knowledge base tag during retrieval, confirm that only documents from the corresponding knowledge base are recalled.
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.