What this category of data looks like
This category of data comes from engineering orders, sand, gravel, cement admixture inspection records, aggregate moisture content test reports, mixing station control system logs, and project laboratory ledgers. Daily documents include mix proportion notices, opening identification forms, production and transport orders, raw material inspection reports, test block curing records, and concrete strength reports. Data updates run per vehicle production batch, per batch of raw material entry, and daily laboratory testing cycles. Each document includes fields such as raw material model, moisture content, mix proportion parameters, slump requirements, and strength grade, with units including kilograms, cubic meters, percentage, megapascals, and others.
What constraints these characteristics impose on knowledge base retrieval and recall
Data is updated frequently per vehicle, per batch, and daily. This requires knowledge base retrieval to support real-time synchronization of the latest raw material inspection and production record data, to avoid calling outdated parameters. Document fields include specialized engineering parameters, requiring matching the accuracy of retrieval keywords, such as recall of terms like aggregate moisture content and slump. Data from multiple sources is scattered across different systems, requiring retrieval to support associated recall across multiple datasets. Meanwhile, a single production batch is linked to many documents, requiring limiting the number of single retrieval results to avoid context overload. Test block strength reports must be linked to corresponding production batches, and retrieval must support filtering by dimensions such as engineering order number and production time, to ensure recalled results match the current audit scenario.
How to configure
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match the recall accuracy of concrete professional terms, filter low-related non-engineering documents |
limit | Top 8–12 entries | Control the number of single retrieval results to avoid exceeding the model's context token limit |
searchMode | mixedRecall | Combine keyword and semantic recall, adapt to multiple types of documents such as mix proportion notices and strength reports |
usingReRank | Enabled | Correct sorting bias in semantic recall, ensure results containing key parameters such as slump and moisture content are displayed first |
datasetSearchUsingExtensionQuery | Enabled | Automatically expand synonymous retrieval for keywords such as raw material models and project numbers, cover documents with different expressions |
collectionIds | Configure according to the dataset ID of the business department | Limit the retrieval scope, only recall valid data from the current mixing station and laboratory |
Three Common Mistakes
- Phenomenon: The application returns empty retrieval results, but the knowledge base test interface can normally recall results. Cause: The
collectionIdsconfiguration for limiting retrieval scope is not set, or the dataset associated with the application is inconsistent with the dataset used in the knowledge base test. - Phenomenon: No document content is displayed on the knowledge base page. Cause: The data synchronization switch is not enabled, or the data update frequency configuration does not match the business update rhythm of per vehicle and per batch, resulting in new documents not being imported into the knowledge base.
- Phenomenon: After upgrading FastGPT, clicking the knowledge base module reports an error, and the version number is displayed abnormally. Cause: The dataset-related configuration files were not fully covered during the upgrade, or the console log was not enabled, making it impossible to directly locate the cause of the error.
How to Verify the Configuration
- Perform a knowledge base test retrieval, enter the aggregate moisture content parameter of the current batch, and check whether the recalled results include the corresponding raw material inspection documents.
- After enabling
usingReRank, initiate a retrieval in the application, compare the sorting results with the test interface, and confirm that the reranking logic takes effect. - Check the dataset synchronization status, confirm that each vehicle production data and daily laboratory test data have completed automatic synchronization.
- View the dataset ID associated with the application, confirm that it is completely consistent with the ID configured in
collectionIds, and no cross-site data is mixed in.
The values given here are common starting points and should be measured against the reader's own samples.