What the data for this category looks like
This category of data originates from customer product models, mold flow or forming simulation files, CAD/CAM source files, machining and measurement records, and records related to trial mold samples. Daily documents include mold design task sheets, DFM reports, part engineering drawings, processing technology cards, mold trial reports, mold modification notices, and mold history logs. Teams update data on a per-project basis, with synchronized updates after each mold trial or modification. Most documents contain structured fields, such as dimensional tolerances on part drawings, defect types and rectification counts in mold trial reports. Some drawing-type files use binary format.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The data sources for this category include CAD/CAM binary files, structured process documents, and mold trial records. Retrieval must support mixed-format content parsing. Teams update data per project, mold trial, or modification. Isolate knowledge base datasets by project to avoid cross-project data interference. Most documents contain structured fields such as dimensional tolerances and defect types. Retrieval must support precise matching of field keywords while retaining the original document's chapter hierarchy to prevent loss of contextual association after splitting. High-frequency mold trial and modification updates require recall results to quickly synchronize the latest rectification records, avoiding use of outdated process or defect data.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Mold design and audit documents have high matching accuracy requirements. Filter out low-correlation non-target content |
limit | Top 6–10 results | Mold audits require reference to multiple process cards and mold trial reports. Too many recall results will exceed the context window limit |
searchMode | mixedRecall | Match structured field keywords (such as dimensional tolerances, material grades) and semantically related process content simultaneously |
usingReRank | Enabled | Documents such as mold trial defects and mold modification notices have high semantic similarity. Reranking improves result sorting accuracy |
datasetSearchUsingExtensionQuery | Enabled | Extend user queries to match project requirements and part parameters in mold design task sheets |
collectionIds | Project-specified dataset IDs | Isolate knowledge base datasets by individual project to avoid irrelevant cross-project data interfering with the audit process |
This configuration applies to FastGPT v4.17.0
Three common mistakes
- Phenomenon: Retrieval results split the content of the same set of documents with chapter titles, leading to broken context and failure to match the complete chapter corresponding to user needs. Cause: The context retention rule during retrieval was not configured correctly, splitting the chapter association of the original document and destroying the complete logical structure of the content.
- Phenomenon: After calling the knowledge base interface to create a training order, no expected task progress identifier is returned, or the interface returns incorrect parameters. Cause: Required fields in the interface request were not filled correctly, or an undefined
trainingTypeoption parameter was used. - Phenomenon: It is impossible to export a complete list of details for all mold-related files uploaded to the knowledge base. Cause: The corresponding data export interface was not called, or the export permission and scope parameters were not configured correctly.
How to confirm the configuration is correct
- Upload a mold part drawing document that includes chapter titles and structured fields, initiate a query containing part size keywords, and check whether the recall results retain the original document's chapter hierarchy and are not split.
- Initiate a query for mold trial defects under the current project, and check whether the recall results only include mold trial reports and mold modification notices for the current project, with no data from other projects mixed in.
- Call the knowledge base interface to create a plain text training set, pass different
trainingTypeparameters, and check whether the task status returned by the interface matches the expected function. - Call the file detail export interface, and check that the returned list includes all uploaded mold-related documents, with no missing or duplicate entries.
The values provided here are common starting points and should be evaluated against samples used in local deployments.