What the data for this category looks like
This category of data primarily comes from engineering orders, sand, gravel, cement, admixture inspection records, aggregate moisture content measurement reports, mixing station control system logs, and project laboratory ledgers. Data updates occur with each vehicle’s production record, each batch of incoming raw material inspection, and daily laboratory ledger updates. Common daily documents include mix proportion notices, startup identification forms, production and transport slips, raw material inspection reports, test block curing records, and concrete strength reports. Fields include aggregate usage per cubic meter of concrete, mixing water adjustment amount, slump index, test block curing age, and similar metrics. Units include kilograms per cubic meter, millimeters, days, and other standard industry units.
Constraints for Knowledge Base Retrieval and Recall
Per-vehicle high-frequency updates require the knowledge base to support real-time or near-real-time index updates. This prevents retrieval of expired raw material inspection or production parameters. Mixed retrieval across multiple document types requires clear field matching rules. For example, mix proportion data must link to the aggregate moisture content of the corresponding engineering order. A large volume of production and inspection data generated in a single day increases vector database scan load during retrieval. This requires limiting the recall scope for each single retrieval. The age field in test block strength reports requires exact matching. This avoids invalid results across different age groups. Drafting exception reports requires fast recall of production parameters and inspection records for the corresponding vehicle. Retrieval delay reduces exception response efficiency.
Configuration Settings
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for concrete parameters, and avoids recalling irrelevant mix proportions or inspection records |
limit | Top 8–12 entries | Balances the completeness of retrieval results and context length, and adapts to the number of fields in process documents and exception reports |
searchMode | mixedRecall | Matches both vector semantics and keywords such as raw material names and project numbers, and adapts to multi-field mixed retrieval requirements |
usingReRank | Enabled | Re-ranks preliminary recall results by relevance, and optimizes the sorting of key parameters when drafting exception reports |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms related to process terms. For example, links "slump" to "workability", and adapts to non-standard term retrieval |
collectionIds | Configured by document type | Only retrieves knowledge base collections corresponding to categories such as mix proportions and inspection reports, and reduces scan scope |
Three Common Mistakes
- Symptom: Knowledge base retrieval takes longer than 5 seconds. Logs show elevated delay in the vector retrieval phase. Cause:
collectionIdsis not configured by document type. This triggers a scan of the full knowledge base dataset, exceeding the retrieval load limit for high-frequency updated ready-mixed concrete production data. - Symptom: Fields appear empty or matching fails when retrieving imported multi-row XLSX tables. Cause: Tables are not split by business scenarios such as mix proportion sheets and inspection reports. Directly importing full tables causes loss of field association relationships during text chunking.
- Symptom: Retrieval results include expired raw material inspection data. Cause: Incremental update rules are not configured, or the update frequency does not match the update rhythm of each vehicle and each batch of raw materials. This leads to recall of outdated inspection parameters.
How to Verify Configuration Correctness
- Run a single retrieval, and confirm that the document type of returned results matches the currently drafted process document or exception report.
- Simulate an exception scenario for retrieval. For example, input "aggregate moisture content is high", and verify that recalled inspection records and production parameters fall within the required business time range.
- Review monitoring data for retrieval duration. Adjust the values of
similarityorlimitto keep duration within the range acceptable for the business. - Import a batch of test XLSX table data, and confirm that fields can be correctly recalled and matched.
The values provided here are common starting points. They should be evaluated against the specific samples relevant to the deployment.