What does the data for this category look like
Data for this category primarily comes from CAD/CAM software, order BOMs, sheet metal inventory systems, laser cutter and bending machine equipment logs, and MES manufacturing execution systems. Daily generated documents include 3D part unfolding drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece inspection records, and process flow cards. Data is updated once for each generated order. Equipment operation data is synchronized daily. Bending coefficients and nesting rules are updated each time a process change is made. Data fields include sheet metal model, thickness, unfolding dimensions, bending angle, process number, and more. Units are mostly millimeters, degrees, and square meters.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
This category’s data covers multi-format documents including engineering drawings, CNC programs, and process cards. Update frequency changes in real time with orders and equipment process shifts. Knowledge base retrieval must adapt to multi-source heterogeneous data parsing and real-time synchronization mechanisms. The multi-variety, small-batch business model leads to rapid growth in document volume alongside orders. Retrieval must accurately match specific fields such as order number, sheet metal model, and process number to avoid irrelevant document recall. Core data such as bending coefficients and nesting rules directly impacts production outcomes. Retrieval and recall accuracy must be strictly controlled. Otherwise, production issues like hole position misalignment or sheet material waste may occur. Retrieval must also support mixed retrieval of multi-format documents, covering full business scenario ledger data collection.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the accuracy requirements for sheet metal drawings and process descriptions, filtering low-relevance recall results |
limit | Top 8 entries | Adapts to on-premises hardware resource limits, controlling total context length within model input constraints |
searchMode | mixedRecall | Balances keyword matching for drawing numbers, order numbers, and other identifiers, and vector semantic matching for process descriptions, adapting to multi-format document parsing |
usingReRank | Enabled | Performs secondary ranking on initial recall results, correcting local bias from vector search and improving recall accuracy of core process data |
datasetSearchUsingExtensionQuery | Enabled | Supplements retrieval of associated fields from order BOMs and equipment logs, covering core ledger data such as bending coefficients and nesting rules |
collectionIds | Only specify the sheet metal cutting and bending knowledge base ID | Limits retrieval scope, avoiding irrelevant documents from other categories and improving retrieval efficiency |
Three Common Misconfigurations
- Retrieval results include non-sheet metal cutting and bending process documents: Failing to specify
collectionIds, with retrieval scope covering the entire platform knowledge base and not limiting to the dedicated business collection. - Retrieval returns insufficient results with low matching accuracy: Setting the
similaritythreshold too high, and not enablingusingReRank, causing valid process documents to be filtered out or not recalled. - Missing bending coefficient and nesting rule fields when generating ledgers: Not enabling
datasetSearchUsingExtensionQuery, failing to recall associated data from order BOMs, leading to empty core ledger fields and triggering production issues such as hole position misalignment or sheet material waste.
How to Verify Correct Configuration
- The FastGPT v4.17.0 backend is accessed, the knowledge base retrieval configuration is reviewed, and confirmation is made that only the sheet metal cutting and bending dedicated collection is selected for
collectionIds. - A test query such as "1.5mm cold-rolled sheet bending coefficient" is submitted, and the
similarityscore of returned results is checked to confirm it falls within the 0.75–0.85 range. - Debug logs are enabled, verification is performed that
usingReRankconducts secondary filtering on initial recall results, and the number of results matches the preset 8-item limit set bylimit. - A latest bending process card is uploaded, retrieval is triggered, confirmation is made that the latest document is recalled normally with no update delay. Deployment resource matching is also verified, ensuring no retrieval lag occurs.
The values provided are common starting points and should be measured against the reader's own samples.