What the data for this category looks like
Data related to inspection and process record audit for metal cutting machining comes from product drawings and 3D models, CAPP/CAM systems, machine tool collected data, tool management systems, MES systems, and measuring 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. These contain fields such as tool life, process dimension tolerances, machine tool spindle speeds and feed rates, with units including millimeters, revolutions per minute, meters per minute, and others. Update frequency is triggered by work orders, workpieces, spot checks, or tool changes, with no fixed cycle.
What constraints do these characteristics impose on the "knowledge base retrieval and recall" link
The data characteristics of this category impose multiple constraints on retrieval and recall: Multi-source heterogeneous data sources require retrieval to cover different types of datasets, to avoid missing key information such as process parameters and inspection records. High-frequency update characteristics require configuring a scheduled incremental synchronization mechanism, to ensure retrieved data matches on-site working conditions. Diverse document structures include plain text operation cards, CNC program code, and drawing metadata, so adaptation to different chunking and retrieval rules is needed. Strict tolerance and process matching requirements mean retrieval must balance vector similarity and precise keyword matching, to avoid recalling irrelevant workpiece data.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Process tolerances and tool parameters for metal cutting machining require precise matching. This interval can filter low-relevance process documents and inspection records |
limit | Top 8–12 results | Individual process documents are lengthy, and too many recall results will exceed the context window and affect subsequent audit logic |
usingReRank | Enabled | This category has a large number of similar inspection records and tool parameters. The re-ranking model can optimize result sorting and improve the recall priority of key information |
datasetSearchUsingExtensionQuery | Enabled | Process information related to work orders is scattered across multiple documents. Extended queries can recall associated operation cards, tool lists and other content |
searchMode | mixedRecall | Balances code keywords for CNC programs and vector features of process parameters, adapting to retrieval needs for multiple document types |
collectionIds | Bind corresponding datasets by workshop work order type | Process and inspection documents for different workpieces are classified independently to avoid mixing inspection records from irrelevant workpieces |
Three common mistakes
- Symptom: No knowledge base selection option appears when calling the tool. Cause: The corresponding metal cutting machining dataset is not bound in the application configuration of FastGPT v4.17.0, or dataset retrieval permissions are not enabled.
- Symptom: The re-ranking model passes testing after deployment, but
usingReRankreturns false for every retrieval result. Cause: The deployment path of the re-ranking model is not correctly associated with FastGPT's re-ranking service, or the input recall result format does not meet the model's input requirements. - Symptom: There is a significant difference between retrieval results from online chat and API calls. Online accuracy is higher while API results have larger deviations. Cause: The API call does not pass the correct context parameters, or the
datasetSearchUsingExtensionQueryconfiguration is not enabled, leading to a retrieval scope inconsistent with the online scenario.
How to confirm the configuration is complete
- Check the application's dataset binding configuration to confirm that the exclusive datasets for process routes, operation cards, inspection records and other content of the corresponding workshop have been associated.
- Initiate a test retrieval, enter a query containing tool wear and dimensional tolerances, and check whether the returned results include documents exclusive to this category, and whether the sorted results after re-ranking conform to business logic.
- Call the API interface, verify that the passed
collectionIdsparameter matches the bound dataset ID, and check whether thesimilarityvalue of the returned results falls within the preset interval. - Simulate work orders for different workpieces to trigger retrieval, confirm that the recall results only include process and inspection documents for the corresponding workpiece, and no data from other workpieces is mixed in.
The values provided here are common starting points and should be evaluated against the reader's own samples.