What this category of data looks like
Data for this category mainly comes from CAD/CAM software, order BOMs, sheet metal inventory systems, laser cutter and bending machine device logs, and MES manufacturing execution systems. Daily documents include part unfolding drawings, CNC cutting programs, bending process cards, nesting drawings, first-piece inspection records, and process flow cards. Data updates trigger per order, daily production shift, or each process change. Document fields include bending compensation coefficient, sheet thickness, hole position tolerance, cutting dimensions, and more. Units are mostly millimeters and kilograms.
What constraints these characteristics impose on knowledge base retrieval and recall
Data sources are scattered across multiple industrial devices and management systems. Cross-source integrated recall logic must adapt to raw data in different formats. High-frequency updates require retrieval links to support incremental synchronization, avoiding performance loss from full refreshes. Most engineering documents are unstructured drawings and program text. Adjust text splitting and vector matching rules to adapt to professional terminology. Retrieval for dedicated fields such as bending coefficients and nesting parameters requires limiting matching ranges to avoid interference from irrelevant results. The multi-variety, small-batch production model leads to a large number of knowledge base entries. Control the number of single-recall results to prevent context overflow that affects generation quality.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Engineering parameter matching accuracy for sheet metal cutting and bending has high requirements. A too-low threshold will introduce irrelevant nesting or bending parameters. A too-high threshold will miss valid compensation value data. This setting adapts to the vector matching logic of FastGPT v4.17.0 |
limit | Top 6 results | The multi-variety, small-batch scenario has a large number of knowledge base entries. Limiting the number of results avoids context overflow, while covering common process parameters and exception handling solutions |
searchMode | mixedRecall | Engineering documents contain both structured parameters and unstructured text descriptions. Hybrid retrieval can match keywords such as "bending interference" and vector semantic features at the same time |
usingReRank | Enabled | Semantic similarity of engineering documents is easily disturbed by irrelevant fields such as drawing numbers and device models. Reranking filters low-correlation results and improves accuracy |
datasetSearchUsingExtensionQuery | Enabled | Queries for process drafting and exception reports often contain omitted engineering terms. Extended queries can supplement implicit keywords such as "bending compensation coefficient" and "nesting surplus material" to improve recall coverage |
collectionIds | Bind to the dedicated knowledge base collection for sheet metal cutting and bending | Isolate knowledge base data from other categories, avoid recalling irrelevant stamping or welding process documents, and narrow the retrieval scope |
Three Common Misconfigurations
- Phenomenon: Retrieval results mix stamping process bending parameter documents, leading to incorrect subsequent process parameter matching. Cause: The
collectionIdsdedicated knowledge base collection restriction is not configured, resulting in cross-category data being recalled. - Phenomenon: Manually uploaded nesting drawing documents cannot be retrieved. Cause: Dedicated text parsing rules for CAD drawing formats are not configured, and
datasetSearchUsingExtensionQueryis not enabled, resulting in failed vector embedding. - Phenomenon: Manually inserted custom index entries disappear after several hours. Cause: Incremental synchronization trigger rules are not configured. Full synchronization overwrites manually inserted non-automatically generated indexes, which is a typical problem of general knowledge base index exceptions.
How to Verify Successful Configuration
- Upload a sheet metal bending process card to the bound dedicated knowledge base collection. Initiate a retrieval containing "bending compensation coefficient". Verify that the recall results only come from this collection, with no documents from other categories mixed in.
- Enter a query containing "bending interference". Verify that the recall results include corresponding exception handling solutions and nesting adjustment parameters, and that the result ranking conforms to semantic relevance.
- Adjust the
similaritythreshold and observe changes in the number of retrieval results. Confirm that the threshold takes effect. - Manually insert a custom process flow card. Wait 1 hour and initiate retrieval again. Verify that the document is not automatically deleted.
The values provided here are common starting points and should be evaluated against your own samples.