What this category of data looks like
Typical data sources include BIM detailed design documents, rebar and concrete batch ledgers, mold table and curing kiln operation records, laboratory test reports, and QR code traceability data. Data updates follow per-single-component, per-batch, or per-kiln schedules. Covered documents include component detailed drawings, rebar cutting lists, mold inspection records, pouring and vibration records, curing records, and demolding and factory inspection reports. Documents include professional fields such as component number, rebar specification, strength grade, mold table number, and curing duration. Common engineering units such as MPa, cubic meters, and days are used.
What constraints these characteristics impose on knowledge base retrieval and recall
Precast concrete component retrieval and recall must meet strict requirements for professional engineering scenarios. First, data updates follow per-single-component, per-batch, or per-kiln schedules. Retrieval must accurately link batches and component numbers to avoid incorrect matches across categories or batches. Second, document types include drawings, tables, ledgers, and other formats. The system must support unified parsing of unstructured and semi-structured data. Third, professional fields such as sleeve anchorage length and demolding strength require the retrieval model to accurately match domain terminology. The system must also associate related documents such as rebar cutting lists and mold inspection records to ensure completeness of bid documents and quotations. Fourth, data has high relevance. Information for a single component is scattered across multiple documents, so the system must support associated recall across documents.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Precast concrete components have dense professional terminology. Filter low-match irrelevant documents to avoid professional errors in bid quotations. |
limit | Top 8 entries | Bid documents must cover the full component list. Excessive retrieval results increase document organization and verification workload. |
usingReRank | Enabled | Component data has strong relevance. The reranking function prioritizes highly relevant documents such as mold inspection and curing records to improve retrieval efficiency. |
datasetSearchUsingExtensionQuery | Enabled | Supplement query conditions for fields such as component number and strength grade to resolve vague references in bid scenarios. |
collectionIds | Specify the exclusive knowledge base ID for precast concrete components | Avoid mixing documents from other categories, ensuring retrieval results only cover the target business scenario. |
Three common errors
- Uploading javadoc-format documents or detailed drawing PDFs results in no retrieval results. Cause: No parsing rules are configured for the corresponding format, so non-standard precast concrete component documents cannot be correctly chunked and embedded.
- GPU resources are not called during knowledge base retrieval. Cause: GPU acceleration configuration for the knowledge base is not enabled, or retrieval parameters do not match GPU-compatible model specifications.
- Retrieval results include documents unrelated to precast concrete components. Cause: Exclusive
collectionIdsare not specified, so the retrieval scope covers knowledge base content from other categories.
How to confirm correct configuration
- Submit a query containing a precast concrete component number, and verify that returned results only include exclusive documents for this category.
- View the FastGPT retrieval log to confirm that parameters such as
similarityandusingReRankare correctly loaded. - Test retrieval results for multi-format documents including detailed drawing PDFs and rebar cutting list Excel files, confirming all can be correctly recalled.
- Enter vague references such as "the curing records of this batch of components" to verify that the extended query function associates documents corresponding to the target batch.
The values provided are common starting points and should be measured against the reader's own samples.