What Data for This Category Looks Like
Data for this category comes from design history files, device and material batch records, MES systems, clean environment monitoring equipment, sterilization equipment, and inspection and calibration systems. Daily documents include production process specifications, batch production records, clean zone operation records, sterilization batch records, inspection reports, software version logs, adverse event and product traceability materials. Data updates occur at the completion of each product batch or equipment unit. Updates also follow clean environment monitoring plans for regular intervals. Document fields include unique device identifiers, material batch numbers, sterilization parameters, clean zone temperature and humidity values, test result values, and more. No unified universal format exists for these documents.
Constraints for Vector Models and Indexes
Multi-source heterogeneous data sources require index systems to support cross-system and cross-format data access and unified encoding. High-frequency, batch-triggered update requirements mean indexes must support incremental updates. This avoids resource consumption from full index rebuilding. Medical-specific compliance and traceability data require indexes to have strict permission control mechanisms. Documents contain specialized medical terminology and specific business fields. Vector models must adapt to medical domain semantic encoding to avoid deviations from general-purpose models. Single data entries vary widely in length. Ranges span short-text test results to long-text process specifications. A flexible chunking strategy is needed to accommodate these different text lengths.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | This category of documents includes long-text process specifications and short-text test records. This range balances semantic completeness and indexing efficiency |
| Overlap Length | 100–150 characters | Prevents loss of key information such as cross-chunk sterilization parameters and traceability identifiers when long texts are chunked |
| Similarity Threshold | 0.75–0.85 | Medical terminology semantic similarity requires a higher threshold to filter irrelevant results, while covering compliance-related similar documents |
| Single Search Limit | Top 10 entries | Bid document and quotation preparation requires reference to multiple batches of compliance records. 10 entries cover typical reference ranges without increasing retrieval time |
defaultToken | 4096 | Adapts to the average length of most medical documents, avoiding truncation of key sterilization parameters and traceability information during chunking |
maxToken | 8192 | Supports complete vectorization of long-text process specifications, meeting the long-context requirements of bid documents |
Three Common Misconfigurations
- Attempting to switch to a vector model other than ada 002 results in an undefined model error. Cause: The access permissions for the corresponding model were not correctly configured in the platform backend, or the entered model name does not match the platform's preset verification format.
- Semantic retrieval returns high scores but results do not match the compliance requirements of bid documents. Cause: The similarity threshold was not adjusted for medical professional documents. General thresholds lead to the recall of a large number of non-critical similar records.
- Configuring one data set to correspond to multiple vector groups results in a save failure. Cause: For FastGPT v4.17.0, multi-vector configuration must be completed during knowledge base creation. Existing knowledge bases cannot have this setting modified. Each vector group must correspond to an independent chunking rule.
How to Verify Successful Configuration
- Upload a typical batch production record document. Check if the chunked text length falls within the preset chunking range. Ensure no obvious truncation of key information occurs.
- Enter a bid-related search term. Review the similarity distribution of retrieval results. Adjust the threshold to a range that meets business requirements.
- Test incremental update of a new batch record. Confirm that the index updates normally and no error messages are displayed.
- Review the vector model call logs. Confirm that the used model name matches the configured settings. No unauthorized call failure records should appear.
The values provided here are common starting points. They should be evaluated against the samples applicable to the intended deployment.