Data characteristics for this category
Daily documents for this category include formula sheets, water treatment records, blending records, sterilization or aseptic filling records, packaging material inspection reports, and finished product inspection reports. Data sources cover water source and raw material testing, formula systems, online sugar and acidity monitoring equipment, filling equipment, visual inspection systems, and LIMS systems. Data is updated per production batch, per shift, or via continuous collection. Individual documents include fields such as batch number, raw material input volume, process parameter values, and inspection results. Field units include concentration units, temperature units, volume units, and other standard measurement units.
Constraints for source citation and traceability
The multi-source scattered data and batch-based document structure of this category require traceability to be tied to the production batch dimension. This prevents mixing data across batches. Continuously collected real-time process data requires traceability to support fast association with real-time data. Production scenarios with switched product lines require traceability to accurately distinguish documents and parameters for different product lines. Exception reports must link to original inspection and process records for the corresponding batch to ensure traceability accuracy. This requires limiting retrieval scope to data sources within a specified production period.
Configuration settings
This page is compatible with FastGPT v4.17.0. The following configuration items are set for this version:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum context length | 800–1200 characters | Matches the average length of single-batch process documents to avoid interference from irrelevant data |
| Single recall limit | Top 6 entries | Covers core process and inspection records within a single batch to avoid redundancy |
| Reranking | Enabled | Prioritizes data sources with higher relevance to the current batch |
quoteMaxToken | 300–500 token | Limits the length of a single citation to ensure clear traceability information referenced in exception reports |
responseData | Return only associated data sources | Filters unassociated knowledge base content to avoid returning irrelevant results |
detail | Enabled | Fully returns fields and source identifiers for cited data to facilitate review |
Three common configuration errors
- Phenomenon: Retrieval results return water treatment records from a non-current production batch. For example, a query for 20240520 batch data returns records from batch 20240519. Cause: Retrieval scope is not limited to the specified batch’s data sources, leading to cross-batch data recall.
- Phenomenon: Cited process parameters do not show complete source identifiers, only parameter values. Cause: The
detailconfiguration is not enabled, so full data source information is not returned. - Phenomenon: A Chinese process query term fails to associate with English equipment calibration records. Cause: Multilingual retrieval adaptation is not configured, or English data sources are not included in the specified retrieval scope.
How to verify correct configuration
- Enter a process query for a specified batch, such as "20240520 batch water treatment records". Confirm retrieval results only include documents from that batch.
- Review the source identifier of cited content. Verify it includes document type, collection time, and batch number.
- Test a query covering multiple data sources, such as "20240520 batch sugar content and sterilization temperature". Confirm recall results include records from the formula system and LIMS system.
- Adjust the
quoteMaxTokenparameter. Confirm the length of a single citation matches expectations.
The values provided here are common starting points and should be evaluated against the reader’s own samples.