What data looks like for this category
This category covers factories producing carbonated drinks, fruit juices, tea beverages, functional drinks, and packaged drinking water. Daily documents include formula sheets, water treatment records, blending records, sterilization or aseptic filling records, packaging material inspections, and finished product inspections. Data comes from water source and raw material testing, formula systems, online sugar and acidity monitoring, filling equipment, visual inspection systems, and LIMS systems. Updates follow per-batch or per-shift cycles, with some physical and chemical parameters collected continuously. Document fields include raw material batch numbers, sugar content values, filling liquid levels, sterilization temperatures, packaging material batch numbers, and others. Units include mg/L, ℃, ml, and others.
Constraints for knowledge base retrieval and recall
This category has multi-source data with widely varying update frequencies. Some parameters are collected continuously. Retrieval must accurately match dimensions such as batch and shift to prevent mixing data across production cycles. Documents include structured physical and chemical indicators and semi-structured inspection records. Retrieval must support weight configuration for different fields to ensure core fields like formula parameters and sterilization parameters are recalled first. Both real-time continuously collected data and batch-updated ledger data exist. The retrieval pipeline must support recall of mixed data types. Since errors directly relate to product safety, recall results must use a relatively high confidence threshold.
Configuration Settings
Based on FastGPT v4.17.0, the following configurations are recommended:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirements for core physical and chemical indicators of this category, such as sugar content and sterilization temperature, to avoid low-confidence results being included |
limit | Top 6 entries | Covers the core record count for a single batch production ledger, while controlling context length |
searchMode | mixedRecall | Balances semantic matching for formulas and inspection records with exact matching for batch numbers and shift numbers |
usingReRank | Enabled | For multi-source mixed production data, re-ranking improves the recall priority of core parameters |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms related to production ledger batches and shifts to match associated fields across multi-source data |
collectionIds | Specified per factory production dataset | Isolates production ledger data from different factories to avoid mixing data across production sites |
Three Common Misconfigurations
- Empty retrieval results or no matching production records:
datasetSearchUsingExtensionQueryis not enabled, so automatic matching of batch and shift related retrieval expansion terms is not possible, leading to failure to recall batch-updated ledger data. - In private deployment scenarios, adjusting the
limitparameter to more than 6 entries does not take effect: The default restriction on retrieval upper limit is not lifted in API configuration, causing requests to be truncated. - Importing production documents prompts a file size limit error: The knowledge base single-file upload threshold is not adjusted to fit the size of production documents. Multi-page finished product inspection records and detailed formula sheets for this category usually exceed the default 1MB limit.
How to Verify Correct Configuration
- Enter a query term that includes a specific production batch, check if the batch field in the returned results matches the query content to verify that the dimension filtering configuration is working.
- Call the API and adjust the
limitparameter to 8, confirm that the number of returned results meets expectations to verify that the retrieval upper limit configuration is working. - Upload a production inspection record larger than 1MB, confirm that the import is successful to verify that the file upload threshold configuration is correct.
- Enter a query term that includes sugar content and sterilization temperature, check if the returned results sort core physical and chemical indicators first to verify that the re-ranking configuration is working.
The values provided are common starting points and should be evaluated against your own samples.