What data for this category looks like
Data related to precast concrete components comes from BIM or detailed design documents, rebar and concrete batch ledgers, mold table and curing kiln operation records, laboratory test reports, and product QR code traceability data. Daily documents include component detailed drawings, rebar cutting lists, mold inspection forms, pouring and vibration records, curing records, and demolding and factory inspection reports.
Data updates are triggered by individual components, each batch of materials, or each curing kiln batch. Full process records cover core steps including component splitting, embedded part verification, and strength control.
What constraints these characteristics impose on knowledge base retrieval and recall
Data updates for this category are triggered by individual components, each batch of materials, or each curing kiln batch. Retrieval and recall must quickly respond to real-time updated process and inspection data.
Document types include drawings, ledgers, test reports and other multi-format content. The system must support cross-format keyword matching and associated retrieval.
Each component corresponds one-to-one with on-site lifting positions. Retrieval must accurately match unique identifiers such as component numbers and batch numbers. This avoids generalized recall of irrelevant content.
Errors in core steps can lead to component scrap or project delays. Retrieval and recall results must have high confidence to reduce invalid or incorrect matching results.
How to set configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting Range | Rationale |
|---|---|---|
similarity | 0.70–0.85 | Precast component process and inspection documents have high matching accuracy requirements, and low-match irrelevant content must be filtered out |
limit | Top 5–10 results | The number of process documents for a single batch of components is limited. Too many results will increase the burden of subsequent screening |
usingReRank | Enabled | When multiple format documents are recalled together, the initial sorting may not align with business priorities. Re-ranking optimizes matching relevance |
searchMode | fullTextRecall | Some process documents contain precise dimensions and position information marked on drawings. Keyword matching ensures accurate recall of core fields |
datasetSearchUsingExtensionQuery | Disabled | Process parameters for precast components are mostly batch-level precise information. Extended queries may introduce irrelevant general content |
Three common configuration errors
- Symptom: Retrieval result order does not change after enabling the re-ranking model. Cause: The call key or model path for the re-ranking model is not configured, or the re-ranking model input format does not match the current knowledge base documents.
- Symptom: Long text entered in the knowledge base is forcibly split, destroying the semantic integrity of key paragraphs. Cause: No rule is set to prevent specified text from being split. The system processes all document content using default chunking rules.
- Symptom: Retrieval results only return preset QA entries, and complete process documents or ledger content cannot be recalled. Cause: The search mode is limited to QA-only matching, and the full document retrieval scope is not covered.
How to verify correct configuration
- Upload a rebar cutting list and curing record for a precast component, perform targeted retrieval, and check whether the matching degree of returned results aligns with the set similarity threshold range.
- Enable the re-ranking model switch, perform multiple retrievals, and observe whether the returned results are sorted in a way that aligns with business logic.
- Enter a long text that requires complete semantic retention, perform retrieval, and confirm that the text is not forcibly split.
- Switch the search mode, and verify whether recall results under different modes match the expected retrieval scope.
The values provided here are common starting points. They should be evaluated against the reader's own samples.