What this category of data looks like
This category of data comes from design history files, component 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 environment records, sterilization batch records, inspection reports, software version records, and adverse event traceability materials. Data is updated upon the production of each batch of products or each piece of equipment. Clean environment monitoring follows a fixed update schedule. Document fields include batch unique identifiers, key material batch numbers, sterilization parameters, clean zone monitoring values, inspection results, and some fields must comply with medical device unique identifier regulations.
Constraints on knowledge base retrieval and recall
This category of data is updated per batch or per piece of equipment, and must be associated with a unique batch identifier. Therefore, knowledge base retrieval must support filtering recall results by batch dimension to avoid mixing data across batches. Document types cover multiple structured and unstructured content categories including process specifications, batch records, and environmental monitoring data. Retrieval must adapt to the field formats of different documents to ensure that key parameters such as sterilization temperature and clean zone colony count are accurately matched. Additionally, product risk levels are high. Recall results must strictly match controlled documents for the current batch, and conflicting content must be avoided to ensure the accuracy of production ledger summarization. Data sources involve multiple systems, so the retrieval workflow must support formats exported from different systems to reduce field mapping errors.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of key production fields such as sterilization parameters and clean zone indicators. A threshold that is too low introduces irrelevant data. A threshold that is too high may miss valid batch records |
limit | Top 8 entries | Production ledger summarization needs to cover multiple document types. Too many results exceed the context window. This category has a limited number of documents per batch. 8 entries cover core process, record, and monitoring data |
searchMode | mixedRecall | This category has structured parameters such as sterilization temperature suitable for keyword matching. It also has process description content suitable for vector recall. The hybrid mode improves recall accuracy |
usingReRank | Enabled | This category requires avoiding conflicting recall content. Reranking sorts results by relevance. It prioritizes documents that best match the current batch’s production requirements |
datasetSearchUsingExtensionQuery | Enabled | This category requires associating batch unique identifiers. Extended queries supplement fields such as batch numbers and medical device unique identifiers. This improves accurate matching of recall results |
collectionIds | Bind to the category-specific knowledge base collection | This avoids recalling production data from other categories. It ensures the retrieval scope only covers controlled documents for medical device assembly and cleanroom production |
Three common mistakes
- Phenomenon: The AI response has an overly high proportion of general knowledge. It does not effectively reference professional knowledge base content, even if retrieval parameters are dynamically configured. Cause: The
collectionIdsorsimilaritythreshold parameters are not passed correctly. This leads to the retrieval scope not covering category-specific documents, and the proportion of professional knowledge base recall results is insufficient. - Phenomenon: Conflicting content with different values for the same sterilization parameter appears in retrieval results. Cause: Filtering recall results by batch unique identifier is not configured. This causes historical data across batches to be recalled simultaneously.
- Phenomenon: Only text datasets are successfully referenced. Structured table data is not recalled. Cause: The retrieval parsing switch for the corresponding structured dataset is not enabled. Or the retrieval configuration does not adapt to the matching rules for table fields.
How to confirm proper configuration
- Upload a single batch’s batch production record and clean zone record for this category. Initiate a query targeting the sterilization parameters of that batch. Verify whether the returned results include the uploaded documents.
- Import two conflicting sterilization batch records. Initiate a query. Verify whether the system returns results sorted by relevance, or if conflicting content can be excluded via filtering conditions.
- Upload a structured table dataset. Initiate a query targeting the fields in the table. Verify whether the returned results include the data content in the table.
- View the retrieval logs in FastGPT v4.17.0. Confirm that configuration parameters such as
searchModeandusingReRankhave taken effect as set, with no missing parameters.
The values provided are common starting points and should be measured against the reader's own samples.