What this category of data includes
This category of data covers production process specifications, batch production records, clean zone environmental monitoring records, sterilization batch records, inspection reports, software version ledgers, and adverse event traceability materials. Data sources include design history files, device and material batch ledgers, MES systems, clean environment monitoring equipment, sterilization equipment, and inspection and calibration systems. Updates run per product batch or per device. Updates are also synchronized regularly per clean environment monitoring plans. Documents include fields such as controlled numbers, batch numbers, sterilization parameters, inspection indicators, and software version numbers. Some documents include aseptic test images and batch association identifiers.
Constraints for knowledge base retrieval and recall
This category of data has strong batch correlation. Retrieval must support filtering by fields such as batch number and clean zone ID. This avoids cross-batch data interfering with bid document accuracy. Some documents include aseptic test images and sterilization parameter tables. Retrieval must support indexing non-text format data. Data updates run at high frequency per production batch or monitoring plan. Recall must support incremental synchronization mechanisms. Special medical device fields such as unique device identification and biocompatibility test results require priority matching during retrieval. This avoids generalized recall of irrelevant content. Retrieval must also adapt to field parsing requirements for structured Excel tables.
Configuration settings
This configuration uses the knowledge base feature of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets precision requirements for matching medical device-specific fields, and avoids recalling low-correlation batch records |
limit | Top 8 entries | Controls context length during bid document preparation, and adapts to content needs for referencing multiple documents |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Reorders recall results by business relevance, and prioritizes displaying core bid documents such as sterilization parameters and clean zone records |
collectionIds | Collection IDs bound per production batch | Enables filtering retrieval results by batch, and avoids cross-batch data interfering with quotation preparation |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands search terms to cover associated terms for process specifications and inspection reports, and adapts to cross-document referencing needs for bid documents |
Three common configuration mistakes
- Scenario: Retrieval results return only text content, not aseptic test images included in documents. Cause: Knowledge base image indexing is not enabled, and only text content is vectorized.
- Scenario: After uploading an Excel table with sterilization parameters, retrieval results cannot match batch number fields. Cause: Dedicated parsing rules for medical device structured tables are not configured, so key fields such as batch numbers are not correctly extracted.
- Scenario: Retrieval results cannot match custom medical device-specific index fields. Cause: Custom indexes for unique device identification and batch numbers are not added, so the system cannot accurately associate business fields.
How to verify correct configuration
- Access the FastGPT knowledge base backend, and check if the
similarityparameter value falls within the 0.75–0.85 range. - Upload an Excel table with sterilization parameters, run a search that includes "batch number", and confirm relevant results are returned.
- Select the collection ID corresponding to the current bid project, run a search, and confirm returned results only include documents from that batch.
- Upload an aseptic test image, run a search that includes "aseptic test", and confirm the document associated with the image is recalled.
The values provided here are common starting points, and should be evaluated against samples relevant to the actual deployment scenario.