Data Characteristics for This Category
Core data for pharmaceutical dosage production comes from approved formulation processes, material batches, MES/electronic batch records, equipment and environmental monitoring data, LIMS, and quality management systems. Daily documents include master production files, batch production records, batch packaging records, cleaning records, inspection records, deviations/CAPA/changes, and batch release review forms. Data updates occur with each production batch, each critical process step, and each completed monitoring plan. Documents cover structured parameters, semi-structured deviation descriptions, and unstructured operation records. Fields include material batch numbers, process parameter values, deviation trigger conditions, release thresholds, and units such as milligrams, degrees Celsius, and batch identification numbers.
Constraints Imposed on Knowledge Base Retrieval and Recall
Multi-source heterogeneous data sources require retrieval to cover cross-system associated information, to avoid missing linked data between material batches and equipment monitoring. Frequently updated production data requires the knowledge base to support incremental synchronization, to ensure retrieval results match the current production batch. Diverse document structures require retrieval to adapt to structured parameters, semi-structured deviation records, and unstructured operation texts, to ensure matching accuracy. Under GMP compliance requirements, recall results must include source identifiers and batch association information, to support subsequent batch release review and data traceability. Additionally, the high severity of potential errors requires retrieval results to have high confidence, to avoid returning irrelevant or incorrect historical data.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Ensures accurate matching of process parameters and deviation descriptions, avoids retrieving irrelevant historical batch data |
limit | Top 8 entries | Balances context length and information coverage, adapts to the multi-field content of batch records |
searchMode | fullTextRecall | Adapts to long-text and structured field retrieval for batch production records |
usingReRank | Enabled | Performs secondary sorting on recall results by GMP compliance and relevance, prioritizes documents related to deviations/CAPA |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval scope to cover associated material batches and equipment monitoring data |
collectionIds | Configured by production department and QA department groups | Limits retrieval scope, avoids mixing historical data not associated with the current batch |
Three Common Configuration Errors
- Issue: Image links in retrieval results display as
input an imageand cannot be displayed normally. Cause: No multi-modal retrieval adaptation rules are configured. The knowledge base only stores image links without associating parsable text descriptions, so the LLM cannot recognize image content. - Issue: After local deployment, knowledge base training is executed, and the data processing step returns empty results. Cause: No chunking rules are configured for the structured format of GMP documents. Long-text batch records cannot be split correctly, so the system skips chunks with no valid content.
- Issue: Retrieved documents are not associated with the corresponding batch number or material information, which cannot support batch release review. Cause: Retrieval scope is not limited via
collectionIds, irrelevant historical data across batches is retrieved, and the field association logic for extended queries is not enabled.
How to Verify Correct Configuration
- Perform a simulated retrieval, input the feeding parameters of a specific batch, and verify that the recall results include the batch production record and material batch information for that batch.
- Upload a cleaning record that includes images, and verify that the LLM can normally parse image links in retrieval results into text descriptions.
- Check the knowledge base data processing logs, confirm that the number of chunks for structured documents meets expectations, and there are no empty chunk records.
- Adjust the
similarityparameter range, verify that the relevance of recall results changes as expected with parameter adjustments, aligned with business requirements.
Note: The configuration values listed above are common starting points. They should be evaluated against the samples of the respective target deployment.