What data for this category looks like
Data for this category primarily comes from product drawings and 3D models, CAPP/CAM systems, machine tool collection terminals, tool management systems, MES systems, and measurement and inspection tool data. Daily documents include process routes, operation cards, tool lists, CNC programs, first-piece inspection records, patrol inspection forms, and measuring tool calibration records. Data updates are triggered by individual work order execution, workpiece machining completion, sampling test results, or tool replacement. Each entry corresponds to parameters, inspection results, and tool status for a specific machining operation. Fields include operation number, tool model, machining parameters, inspection values, and more.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The scattered sources and tightly linked nature of data for this category require retrieval to accurately match multiple types of associated documents corresponding to an operation, to avoid mixing data across work orders. The high-frequency update feature requires the retrieval link to support incremental recall logic, to ensure that returned tool lists, inspection records, and other content are in the latest state. Documents contain mixed-format content such as CNC program text and inspection values, so recall rules for structured parameters and unstructured text need to be adapted. The characteristic of concentrated data volume for a single work order requires limiting the number of recall entries per request, to avoid context overload. The strong binding relationship between tools and operations requires retrieval to associate entity dimensions; relying only on keyword matching cannot cover entity association requirements.
How to set configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Parameters and operations in metal cutting machining are closely linked, so low-match irrelevant documents need to be filtered to avoid mixing data from other work orders |
limit | Top 8–12 entries | The number of associated documents for a single work order is concentrated, so limiting recall entries avoids context overflow while covering core process and inspection data |
searchMode | mixedRecall | Need to match both parameter text in CNC programs and structured fields in operation cards, balancing accuracy and recall scope |
usingReRank | Enabled | Professional terminology is common in machining documents; re-ranking can further filter recall results with semantic deviations, improving the accuracy of report generation |
datasetSearchUsingExtensionQuery | Enabled | Supplements implicitly associated data such as tools and measuring tool calibration related to the operation via extended queries, covering full-chain ledger content |
collectionIds | Bind to dedicated knowledge base collections for the corresponding workshop | Isolates work order data from different workshops, avoids mixing documents across workshops, and ensures exclusive accuracy of ledger data |
Three common misconfigurations
- Phenomenon: Retrieval results mix operation card documents from other workshops, and the number of recall entries for a single batch exceeds 15. Cause: The
collectionIdsparameter is not configured to bind the dedicated workshop knowledge base, and thelimitparameter is not set to restrict the recall upper limit, resulting in redundant data being pulled across scopes. - Phenomenon: Tool calibration certificate images in the knowledge base cannot be associated with the retrieval results of the corresponding operation, and only the text description below the image is extracted. Cause: The semantic association rules for drawings and certificate images in metal cutting machining documents are not adapted; only plain text fragments are processed, and key information such as calibration values carried by images is not covered.
- Phenomenon: Retrieval returns no results, or a timeout error is triggered. Cause: The
similaritythreshold is set too high, resulting in insufficient eligible recall entries, or incremental recall is not enabled, causing too many expired machine tool collection data to be loaded.
How to confirm configurations are correctly set
- Enter the knowledge base management interface, check whether the bound
collectionIdscorrespond to the dedicated collection of the target machining workshop, and confirm that no redundant collections from other workshops are associated. - Initiate a retrieval test for a single operation, check whether the number of returned entries matches the preset range, and no documents from unrelated work orders are mixed in.
- Upload a test document containing process drawings and text, initiate a retrieval, and confirm that the results include parameter information and text content associated with the image.
- Simulate a document update after a tool replacement, trigger knowledge base incremental synchronization, initiate a retrieval for the corresponding operation, and confirm that the returned tool list is the latest version.
The values provided here are common starting points, and should be evaluated against the reader's own samples.