What this category of data looks like
Data subjects include commercial concrete mixing plants, project laboratories, and quality scheduling departments. Daily documents include mix proportion notices, opening identifications, production and transport orders, raw material inspection records, test block retention records, and strength test reports. Data sources cover engineering orders, sand, gravel, cement, admixture inspection data, aggregate moisture content measurement results, mixing plant control system logs, and laboratory test data. Update frequency follows per production batch, per raw material delivery, and daily summary. Document fields include raw material proportion parameters, moisture content values, transport time, test block strength grade, pouring location, and other fields. Units include kilograms, cubic meters, percentage, megapascals, and other units.
Constraints on knowledge base retrieval and recall
Frequently updated production and inspection data requires the knowledge base to support incremental synchronization. This prevents recall of expired proportion or inspection records. Multiple types of segmented documents require precise filtering by document type. For example, only recall mix proportion notices for proportion correction. Strong correlations exist between data points. For example, moisture content data must link to water consumption calculations. Retrieval must support cross-document associated recall. Single batch data volume is limited but updates occur frequently. This requires controlling the context length of a single recall to avoid exceeding model capacity limits. Severe error consequences require improving retrieval accuracy. Filter low-relevance or incorrect historical data to ensure accuracy of production and quality processes.
Configuration recommendations
The following configurations are based on FastGPT v4.17.0, and are optimized for the ledger retrieval scenario of ready-mixed concrete production:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches high-precision business scenarios such as ready-mixed concrete proportioning and strength reports, avoiding recall of low-relevance raw material inspection data |
limit | Top 6 entries | The number of documents associated with a single batch production ledger is limited. Excessive recall causes context overload and affects report generation efficiency |
searchMode | embedding | For unstructured test reports and proportion sheets, semantic search accurately matches keywords and business logic, outperforming keyword matching |
usingReRank | Enabled | After multi-document recall, re-ranking filters redundant records of the same type, such as repeated moisture content measurement data, improving report accuracy |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands business queries. For example, expand "correct mix proportion" to "mix proportion notice combining aggregate moisture content to adjust water consumption" to match more accurate historical data |
collectionIds | Bind corresponding knowledge base collections by business department | Distinguish inspection records from the quality department and transport orders from the scheduling department, avoiding cross-department data confusion and improving retrieval accuracy |
Three common mistakes
- When running a workflow, custom knowledge base selection variables return null values, making it impossible to specify the retrieval dataset. The cause is that the variable data type is not configured as a knowledge base collection type, or the corresponding business knowledge base is not bound to the optional range of the variable.
- Retrieval results return a large number of duplicate test block strength records, leading to redundant and inaccurate generated reports. The cause is that the
usingReRankparameter is not enabled, and no re-ranking filtering is performed on recall results, retaining duplicate detection data from the same batch. - Knowledge base parsed paragraphs are too long, exceeding model context limits and causing execution failures. The cause is that the maximum paragraph depth for knowledge base parsing is not set reasonably, splitting a single opening identification document into overly long paragraph fragments, increasing context load.
How to verify correct configuration
- Enter the workflow variable configuration page, check that the data type of the custom knowledge base selection variable is set to knowledge base collection type, and the optional range includes the datasets of the corresponding business department.
- Run a test case, input "correct the mix proportion corresponding to today's aggregate moisture content", check that the
similarityscore of the retrieval results falls within the 0.75–0.85 range, and the recalled document types meet expectations. - View the workflow logs, confirm that the number of recalled documents per single retrieval does not exceed 6, and there are no duplicate or expired records.
- Verify that
datasetSearchUsingExtensionQueryis effective. Input a simplified query such as "pouring strength report", check whether it is automatically expanded to a complete query matching the business scenario.
The values provided here are common starting points and should be measured against the reader's own samples.