What Data Looks Like for This Category
The daily documents for this category include unfolding drawings, CNC cutting programs, bending process cards, nesting drawings, first-piece records, and process flow cards. Data sources cover CAD/CAM software, order BOMs, sheet metal inventory systems, laser cutters, bending machines, and MES systems. Data updates are triggered per order, daily, or each time a tooling change is made. Fields vary significantly across different documents: unfolding drawings include bending compensation coefficients and hole position coordinates, CNC programs include cutting path parameters, and process cards specify bending sequences and tolerance requirements.
Constraints Imposed on Knowledge Base Retrieval and Recall
The multi-variety, small-batch business model leads to rapid growth in document volume alongside orders. Retrieval must support precise matching of process parameters and filter overly generalized matching results. Data comes from multiple systems with inconsistent formats, so retrieval needs to adapt to field rules across different data sources. Data updates triggered by each tooling change require near-real-time data source synchronization. Core error-prone steps involve bending compensation and hole position misalignment, so recalled results must prioritize documents linked to process parameters, filter documents with generic descriptions, and avoid irrelevant documents interfering with quality analysis.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precision requirement for matching sheet metal process parameters, filters irrelevant process documents |
limit | Top 8 entries | Adapts to the multi-variety, small-batch document scale of this category, controls context token usage |
searchMode | mixedRecall | Balances semantic matching and exact keyword matching, meets retrieval needs for drawing numbers and bending coefficients |
usingReRank | Enabled | Performs secondary ranking on recalled results, prioritizes documents related to bending compensation and hole position misalignment |
datasetSearchUsingExtensionQuery | Enabled | Optimizes user-input process questions to match professional terminology in unfolding drawings and nesting drawings |
collectionIds | Bound to exclusive sheet metal cutting and bending dataset ID | Isolates category-specific data to avoid recalling documents from non-target processes |
Three Common Misconfigurations
- Phenomenon: The document directory is automatically deleted after the knowledge base is imported, and duplicate content is removed. Cause: The automatic document cleaning directory removal switch was not disabled, and duplicate content deduplication configuration was enabled.
- Phenomenon: The question-and-answer pair extraction task gets stuck in training for a long time, with no related call records in the call log. Cause: Data source access permissions were not configured, or the imported CAD document format was not processed by the system's supported parser.
- Phenomenon: The knowledge base dataset in the development environment cannot be downloaded after export, with a server error prompt. Cause: Storage path permissions for exported files were not configured, or exported file size exceeded the system limit.
How to Verify Correct Configuration
- In FastGPT v4.17.0, upload a sheet metal bending process card, run a search for the query containing "bending coefficient", and check whether the relevance of returned results matches the set
similaritythreshold. - Launch a question-and-answer pair extraction task, review the task log for parsing progress updates, and confirm data source permission configurations are correct.
- Export the knowledge base dataset in the development environment, verify that the download function operates normally, and confirm storage path permission configurations are correct.
- Compare retrieval results with and without
usingReRankenabled, confirm the reranking function takes effect and prioritizes documents related to core process parameters.
The values provided are common starting points and should be evaluated against the reader's own samples.