What Data for This Category Looks Like
Data for this category covers full production process documents for single-use consumables, implantable and interventional devices, diagnostic equipment, and active medical devices. Documents include production process specifications, batch production records, cleanroom environment records, sterilization batch records, inspection reports, software version records, and adverse event traceability materials. Data sources cover design history files, device and material batch ledgers, MES systems, clean environment monitoring equipment, sterilization equipment, and inspection and calibration systems. Updates occur after each production batch or single device is completed. Cleanroom-related data is also updated per a fixed environmental monitoring plan. Data includes structured fields such as batch numbers, unique device identifiers, environmental monitoring parameters, and test results, plus unstructured document content.
Constraints Imposed on Knowledge Base Retrieval and Recall
The multi-source mixed data structure and strongly associated attributes of this category impose clear constraints on knowledge base retrieval and recall. Systems must support filtering recall results using precise fields such as batch numbers and unique device identifiers. This prevents matching data from unrelated production batches. Mixed data sources of structured ledgers and unstructured documents require compatible retrieval rules for different formats. This ensures both types of data can be effectively retrieved. Data updated per batch or fixed environmental monitoring plans requires an incremental sync mechanism. This avoids reloading expired data during full retrieval. Adverse event traceability materials must be strongly bound to their corresponding production batches. Retrieval must automatically associate and recall full-link production and inspection documents for that batch. This ensures the completeness and accuracy of traceability processes.
Configuration Settings
FastGPT v4.17.0 supports the following configurations:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches strongly associated fields such as medical device batches and unique identifiers. Filters low-confidence results to ensure traceability accuracy |
limit | Top 10 entries | Covers multiple types of documents related to a single production batch. Controls context length to avoid exceeding business response thresholds |
searchMode | mixedRecall | Adapts to mixed data sources of structured ledgers and unstructured documents. Balances precise field matching and semantic association recall |
usingReRank | Enabled | Improves the accuracy of association ranking between adverse event traceability materials and production batches. Meets retrieval requirements for high-risk scenarios |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands search terms for fields such as batch numbers and unique device identifiers. Covers multi-source associated production and inspection data |
collectionIds | Specify dedicated dataset IDs for clean production and quality traceability | Isolates data from other categories. Prevents retrieval of unrelated documents. Narrows retrieval scope to improve efficiency |
Three Common Configuration Mistakes
- The symptom is garbled file names or failed file parsing when uploading knowledge base files with Chinese names via API. The cause is incorrect configuration of character encoding parameters in the API request header, which prevents correct parsing of Chinese file names.
- The symptom is that the number of knowledge base retrieval results does not match expectations, and only a small number of related documents are returned. The cause is incorrect setting of the
limitparameter value, or failure to enabledatasetSearchUsingExtensionQuery, which leads to failure to cover associated field retrieval. - The symptom is excessively long single retrieval times and delayed conversation responses. The cause is failure to enable
usingReRankand restrict thelimitparameter value, which leads to loading too many unrelated documents into the context and increasing processing load.
How to Verify Correct Configuration
- Upload a test file with a Chinese name to the specified dataset. Confirm the file name displays normally in the knowledge base list without garbled characters.
- Enter a known batch number or unique device identifier as a search term. Verify that returned results include related documents such as production records and inspection reports for that batch.
- Perform multiple retrieval operations. Check response times. Adjust the
limitorusingReRankconfiguration to optimize retrieval speed. - Check the dataset scope of retrieval results. Confirm that unrelated documents from other categories are not recalled.
The values provided are common starting points and should be measured against the reader's own samples.