What the Data for This Category Looks Like
Data for this category comes from approved formulation processes, MES electronic batch records, LIMS, quality management systems, and equipment environmental monitoring data. Daily documents include master production files, batch production records, batch packaging records, cleaning records, inspection records, deviation CAPA changes, and batch release review forms. Data is updated per production batch, key process steps, and monitoring plans. Fields include material batch numbers, process parameters, test results, deviation records, and more. All data must be stored synchronously with documents required by GMP.
What Constraints These Characteristics Impose on the Workflow Orchestration Link
Data for this category is tightly bound to production batches and must comply with GMP synchronization requirements. Workflow orchestration must therefore trigger nodes according to production batches. Multi-source data sources require workflows to support cross-system data pulling. Material batch numbers, process parameters, and similar fields must be passed as core variables. Batch record documents have many fields and are closely linked. Workflows must configure data linkage between nodes to ensure review and audit links can call full data for the corresponding batch. Since errors directly affect batch release and regulatory compliance, workflows must add verification nodes to block non-compliant data flows.
How to Configure the Settings
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Basis for This Approach |
|---|---|---|
variables | Bind material batch number and production order number fields | Match the core requirement of batch-based data management for this category, ensure context passed between nodes is bound to the corresponding production batch |
| Question Classification | Classify by test results, deviation types, and release status | Cover core judgment scenarios in batch record review, quickly divert different types of audit tasks |
| Loop Settings | Loop by the number of review links per single batch, set maximum loop times to 3 | Match time-consuming scenarios of step-by-step review such as weighing and feeding, avoid infinite loops |
| Maximum Context | Set based on actual measurement | Adapt to the associated call requirements of multi-source data such as LIMS and QMS, avoid context overflow |
| Trigger Condition | Trigger after the completion of key production steps per batch, set delay to 60 seconds | Match the rhythm of data updated by process steps, ensure the workflow starts promptly at the corresponding node |
| File Upload Configuration | Directly pass original file format, set maximum size to 1000 MB | Meet the requirements of storing original documents under GMP, avoid losing metadata after parsing |
Three Common Mistakes
- Phenomenon: The output of the question classification node is empty, or the classification result does not match expectations. Cause: Core fields such as test results and deviation descriptions in batch records are not used as classification basis, only general text is passed in.
- Phenomenon: The number of context items set for the AI node does not match the actual number of called context items, and only a small amount of context is displayed in the conversation details. Cause: Full data context of the corresponding batch is not associated via
nodeId, only temporary context of the current node is called. - Phenomenon: Uploaded files are not correctly recognized by the large model, or are only parsed as plain text with metadata required by GMP lost. Cause: The option to directly pass the original file format is not configured, and the system default text parsing process is used.
How to Confirm the Configuration Is Correct
- Trigger the production node corresponding to a batch, check whether the workflow starts according to the preset rhythm, with no delays or omissions.
- Call the question classification node, pass in the test record text of a single batch, verify that the classification result matches the preset rules.
- Start the loop node, simulate the review links of a single batch, confirm that the number of loops matches the number of links, with no abnormal termination.
- Upload the original batch record file, check whether the large model can directly read the file content without losing metadata.
The values given are common starting points and should be measured against the reader's own samples.