What data for this category looks like
Data sources include original equipment manufacturer (OEM) full machine manuals, parts BOM lists, on-site operation and maintenance archived logs, and annual maintenance specification documents. Sources are internal archives of central and local state-owned enterprise equipment manufacturing firms and OEM supplier materials.
Update cadence: OEM materials are updated every 6 months with model iterations. Internal operation and maintenance documents are adjusted monthly based on on-site issues.
Document structure is mostly chaptered technical texts. It includes technical parameters with units such as full machine weight, rated power, and hydraulic system pressure. It also includes step-by-step process content such as full machine on-site installation, troubleshooting, and warranty processing.
What constraints these characteristics impose on knowledge base retrieval and recall
Precise technical parameters with units require full matching logic during retrieval, to avoid matching failures caused by splitting units.
Step-by-step process documents require retaining contextual coherence during recall. Isolated step fragments must not be retrieved alone.
Data source formats include PDF, Word and scanned documents. Unified chunking processing is needed to retain paragraph structure.
Update frequency is relatively stable, but individual document length is long. Chunk length must balance semantic completeness and retrieval efficiency.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Matches documents with precise parameters for construction machinery. A threshold that is too low will include irrelevant technical content, while a threshold that is too high will miss valid process steps |
limit | Top 6 results | Individual construction machinery documents have long content. Too many recalled results will exceed the context window, while too few will fail to cover the full troubleshooting process |
searchMode | semantic | For documents with technical parameters and process steps, semantic retrieval better matches users' vague fault descriptions than keyword retrieval |
usingReRank | Enabled | Construction machinery documents have many technical details. Re-ranking filters redundant results that are semantically similar but do not match parameters |
datasetSearchUsingExtensionQuery | Enabled | For users who describe equipment issues in colloquial language, extended queries cover professional terms in OEM manuals |
collectionIds | Fill in the bound construction machinery knowledge base collection ID | Accurately limit the retrieval scope to avoid mixing document data from other categories |
Three common misconfigurations
- Symptom: Semantic retrieval returns score values outside the 0-1 range, reaching thousands. Cause: The normalization logic for the
similarityparameter is not configured correctly, resulting in raw similarity scores not being scaled. - Symptom: No source file information for the retrieval is obtained after API calls. Cause: The knowledge base's source file traceability configuration is not enabled, or the associated field for returning source files is not specified in the call parameters.
- Symptom: Retrieval results are empty in some scenarios. Cause:
searchModeis incorrectly set to pure keyword mode, which cannot match vague technical descriptions in construction machinery documents, or thesimilaritythreshold is set too high, filtering out valid results.
How to confirm the configuration is correct
- Initiate a query that includes specific equipment parameters, and check if the
similarityscores of returned results fall within the preset range. - View the source file list of retrieval results, and confirm that only construction machinery-related knowledge base documents are included.
- Initiate a query with a vague fault description, and check if the corresponding troubleshooting process documents are returned.
- Adjust the
similaritythreshold, and observe whether the number of recalled results changes as expected.
The values provided are common starting points and should be evaluated against samples specific to the deployment context.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.