What the data for this category looks like
The core data for this category comes from orders and size charts, CAD patterns, fabric and accessory inspection records, hanging system process data, needle detector test results, and customer sample-sealed documents. Daily maintained documents include process sheets, pattern drawings, layout plans, cutting lists, first-piece sample-sealed records, process quality sheets, and packaging size-matching lists. Data is updated per order, per style, per batch. Documents for a single-style order include fields such as size number, fabric color number, cut piece quantity, process parameters, with units including centimeters, piece count, batch number, and others.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The data for this category has a high proportion of semi-structured content and a high update frequency. This requires the retrieval link to support incremental fast updates. The irreversible nature of cut pieces means retrieval results must accurately match parameters such as the current order’s size and fabric color number, to avoid batch rework caused by mismatches. The complex process disassembly feature requires retrieval to quickly recall corresponding process requirement documents using process keywords. Cross-batch data must also be strictly isolated, to prevent documents from old-style orders from being mixed into the current batch’s retrieval results.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Garment cutting and sewing has high requirements for matching accuracy. This range filters low-matching irrelevant documents and avoids process mismatches |
limit | Top 8 entries | Single-style orders have a large number of process documents. This value covers core process requirements while controlling context length |
usingReRank | Enabled | Further filters results with insufficient matching accuracy, especially for precise fields such as fabric color number and cut piece number |
datasetSearchUsingExtensionQuery | Enabled | Converts user-input process descriptions into query terms matching document fields, adapting to the retrieval needs of semi-structured documents |
searchMode | mixedRecall | Balances precise matching of structured fields such as size and color number, and semantic matching of process descriptions |
Three common mistakes
- Phenomenon: After local deployment, uploading process sheets, layout plans and other files, the second-step data processing returns empty results. Cause: No adapted parsing rules are configured for semi-structured documents of the garment category, such as vector format CAD patterns, causing files to fail correct splitting and import into the knowledge base.
- Phenomenon: When calling the knowledge base retrieval interface, the number of returned results is unstable, sometimes with no matching results. Cause: The
similaritythreshold is not fixed, or thelimitparameter is not set to a clear value, causing retrieval results to change with minor fluctuations in the query vector. - Phenomenon: Retrieval results include old-style cut piece data unrelated to the current order. Cause: The current order’s corresponding knowledge base collection is not limited via
collectionIds, causing cross-batch historical data to be retrieved.
How to confirm the configuration is correct
- Upload the process sheet and layout plan for a new order, wait for FastGPT v4.17.0 to complete data processing, then check the knowledge base file list to confirm successful import.
- Enter the test query "XX size top garment cutting process requirements", verify that retrieval results include the process sheet document for the corresponding order, and that the
similarityscore falls within the preset 0.75–0.85 range. - After enabling
usingReRank, compare retrieval results before and after enabling to confirm that irrelevant documents with low matching accuracy are effectively filtered. - Conduct a batch incremental update test, upload packaging size-matching lists for a new batch, verify that new files can be quickly retrieved with no data loss or duplicate import.
The values provided are common starting points and should be measured against the reader's own samples.