What this category of data looks like
Production data from precast concrete component factories covers PC components, pipe piles, bridge components and other categories. Daily documents include component detailed drawings, steel bar cutting lists, mold inspection records, pouring and vibration logs, curing ledgers, and demolding factory inspection reports. Data sources include BIM detailed design files, steel bar and concrete batch documents, mold table and curing kiln operation data, laboratory test records, and QR code traceability information. Updates are triggered per single component, per batch of materials, and per curing cycle of a kiln. Fields include component ID, steel bar batch number, curing duration, compressive strength value and more. Units include MPa, hours, pieces and others.
What constraints these characteristics impose on knowledge base retrieval and recall
This category of data has high update frequency triggered per single item or batch, requiring the retrieval link to support incremental updates and batch filtering. Multi-source and heterogeneous document types include both structured batch fields and unstructured drawing text, requiring retrieval to balance semantic matching and keyword matching. Data association accuracy requirements are high: incorrect recall will cause serious production problems such as embedded part misalignment and insufficient strength, so retrieval thresholds must be strictly controlled. Ledger summary requires aggregation of multi-source data, requiring retrieval results to be associated and spliced by component ID to avoid data confusion.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Semantic relevance of associated documents for precast concrete components is strong. A threshold that is too low will introduce irrelevant mold inspection records from the same workshop, while a threshold that is too high will miss steel bar cutting lists from the same batch |
limit | Top 10 entries | Associated documents for a single component typically do not exceed 8. Limiting the number of recalled entries avoids context overflow and adapts to the context length requirements of ledger summary |
searchMode | mixedRecall | Balances matching of structured fields such as batch numbers and mold table numbers with semantic retrieval of unstructured drawing text, improving recall accuracy |
usingReRank | Enabled | Semantic similarity of documents related to components is relatively high. Reranking can filter out non-core records with low relevance, optimizing the content purity of report generation |
datasetSearchUsingExtensionQuery | Enabled | User queries often omit component IDs. Extended queries can automatically supplement associated fields such as batches and mold tables, improving recall rate for cross-field retrieval |
collectionIds | Bind the exclusive knowledge base ID for precast concrete component ledgers | Isolates documents from other categories, avoiding recall of irrelevant data from ready-mixed concrete and other construction products, ensuring precise retrieval scope |
Three Common Misconfigurations
- Phenomenon: Generated production reports include curing records from non-target batches. Cause: Failed to configure
collectionIdsto bind the exclusive knowledge base, resulting in recall of production data from other factories or other categories. - Phenomenon: Imported component detailed drawings only extract plain text, and embedded part positioning information marked on the drawings is not retrieved. Cause: Did not enable structured extraction configuration for image content, only processed text fragments within the document.
- Phenomenon: Number of retrieval results is far lower than expected. Cause: Did not adjust the
similaritythreshold; an overly high threshold filtered out some steel bar cutting lists that meet the required relevance standards.
How to Verify Proper Configuration
- Submit a query with a specified component ID, check that recalled documents only include steel bar cutting lists and curing records corresponding to that ID, with no cross-batch data.
- Upload a component detailed drawing with embedded part markings, confirm that retrieval results include positioning text information from the drawing, verifying the extraction effect of image content.
- Test a query without specifying a batch, confirm that
datasetSearchUsingExtensionQueryautomatically supplements associated mold table numbers and concrete batch numbers, verifying that extended queries take effect. - In the knowledge base management interface of FastGPT v4.17.0, view the incremental update log, confirm that documents for each batch of components can be automatically synchronized according to the update frequency.
The values provided here are common starting points and should be measured against the reader's own samples.