Data Characteristics for This Category
Data related to precast concrete components comes from BIM detailed design documents, steel bar and concrete batch ledgers, mold table and curing kiln operation logs, laboratory inspection reports, and QR code traceability systems. Daily documents include component detailed drawings, steel bar cutting lists, mold inspection records, pouring and vibration records, curing records, and demolding and factory inspection reports.
Data updates trigger per single component, each batch of materials, and each curing kiln process. Fields include component number, steel bar specification, pouring strength, curing duration, and more. Units include millimeters, megapascals, hours, and other common engineering units. The character count of individual documents ranges from hundreds to thousands.
Constraints Imposed on Knowledge Base Retrieval and Recall
Precast concrete component data comes from scattered sources, with update frequency tied to production processes. This requires knowledge base retrieval to support mixed recall of cross-system documents, while adapting to incremental update logic to avoid retrieving expired batch or curing data.
Documents include multiple types such as detailed drawings, cutting lists, and numerical records. When chunking content, balance the structured fields of engineering drawings and the splitting rules for text descriptions.
Each single component corresponds to a unique traceability identifier. Retrieval must support precise matching by component number, filter irrelevant results across batches, ensure recalled results are strongly associated with the current audit object, and improve audit efficiency and accuracy.
Configuration Settings
This configuration is formulated for FastGPT v4.17.0, adapted for the inspection and process record audit scenario for precast concrete components.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Inspection records for precast concrete components contain numerous numerical fields and process descriptions. A threshold that is too low will introduce irrelevant batch data, while a threshold that is too high will fail to retrieve valid matching items. Calibrate based on actual testing |
limit | Top 8 entries | A single audit material involves no more than 5 associated component and batch data items. Retaining 8 entries covers relevant information while avoiding context overload |
usingReRank | Enabled | Inspection records contain numerous similar numerical values and process descriptions. Reranking filters low-correlation similar results and improves audit accuracy |
datasetSearchUsingExtensionQuery | Enabled | Component numbers can be associated with corresponding detailed drawings, cutting lists and other documents through extended queries, improving the completeness of cross-document recall |
searchMode | embedding | Need to match unstructured text such as process descriptions and numerical requirements, while supporting precise matching logic by number |
collectionIds | Specify the knowledge base ID according to the audit scenario | Limit the retrieval scope to only documents related to precast concrete components from the current factory, avoiding mixing in knowledge base data from unrelated categories |
Three Common Misconfigurations
- The reranking model passes testing after deployment, but every retrieval result shows rerank as false. The symptom is that retrieval results are not sorted by relevance. The cause is that the
usingReRankparameter is not correctly associated with the deployed reranking model, or the model call link is configured incorrectly. - There is a large difference between online chat and API call results. The symptom is that the recalled results returned by the API are inconsistent with those on the online interface. The cause is that the correct
collectionIdsparameter is not passed during API calls, or thedatasetSearchUsingExtensionQueryconfiguration is not enabled, resulting in the knowledge base scope not being correctly restricted. - Unable to select a specified knowledge base during tool calling. The symptom is that retrieval results include irrelevant document sets. The cause is that the
collectionIdsparameter is not configured to limit the retrieval scope, resulting in the default retrieval of all knowledge bases during calls.
How to Verify the Configuration Is Properly Set
- Upload the detailed drawing and inspection record of a single precast concrete component, initiate a retrieval, and check whether the similarity scores of the returned results fall within the configured range.
- Initiate a retrieval request with a component number, and check whether the recalled results include associated documents such as cutting lists and curing records corresponding to that number.
- After enabling the reranking function, compare the sorting of retrieval results before and after enabling to confirm that highly relevant results appear at the top.
- Use the same input and knowledge base to call the API and the online chat interface, and check whether the returned recalled results are consistent.
The values listed above are common starting points. Adjust them based on your own specific samples and use cases.