What Data for This Category Looks Like
The data for this category typically comes from orders and size charts, CAD pattern files, hanging production systems, needle detector data, and customer sealed samples from garment OEM/ODM workshops. Daily documents include process sheets, layout plans, cutting lists, first article sealed samples, process quality sheets, and packaging size matching lists. Data updates occur per order, per style, and per batch. Each individual document contains fields such as size, number of cut pieces, batch number, and process requirements. Common units are centimeters, pieces, and meters.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link?
This category has a large number of styles and strong seasonality. Cut pieces are irreversible, so retrieval results must accurately match the process requirements of the current order to avoid batch deviations. Data updates happen frequently, so the system must support fast index updates by order and style. This prevents old data from interfering with audits. Documents include unstructured content such as CAD layout plans, so retrieval logic for non-text fields must be supported. Processes are split in detail, so retrieval must support precise matching by detailed fields like process number and size. This ensures accuracy in inspection and audit links.
How to Configure the Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Garment cutting documents require precise matching of process requirements to prevent low-similarity irrelevant layout plans or process sheets from being included in results. Adjusted based on parameter rules for FastGPT v4.17.0 |
limit | Top 8 entries | This category has many processes per order, so enough associated documents must be retrieved. At the same time, context length must be controlled to avoid exceeding system processing limits |
searchMode | mixedRecall | Balances text keyword matching (such as process numbers) and semantic matching (such as process requirements) to adapt to retrieval needs for multiple types of documents |
usingReRank | Enabled | Performs secondary sorting on retrieved results to filter out content that does not match the current order's size and batch, improving retrieval accuracy |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms to cover abbreviated terms in process sheets, improving matching success rates for non-standard terms |
collectionIds | Specified by the workshop to which the order belongs | Limits the retrieval scope to only the process and layout documents for the current batch to prevent cross-order data from interfering with audits |
Three Common Misconfigurations
- Duplicate cutting lists appear in retrieval results, and the order of custom-split document indexes does not match the retrieved results. This occurs because deduplication logic is not configured correctly, causing duplicate documents to be retrieved simultaneously and disrupting the original index order.
- A request timeout error is returned during retrieval, and target process sheets cannot be obtained. This occurs because the
collectionIdsretrieval scope is not limited, and all historical documents from the workshop are retrieved at once, exceeding system processing thresholds. - Layout plans from other batches are included in retrieved results, leading to size deviations during audits. This occurs because the
usingReRankparameter is not enabled, and secondary filtering of retrieved results using order-specific fields is not performed.
How to Confirm the Configuration Is Correct
- Upload a test cutting list and corresponding layout plan, initiate a retrieval, and check whether the target document is included in the retrieved results, and duplicate documents appear only once.
- View the retrieval log to confirm that the number of returned results matches the value set by the
limitparameter. - Manually adjust the value range of the
similarityparameter, observe changes in the number of retrieved results, and verify that the parameter takes effect. - Switch test data from different orders, confirm that retrieval results only include relevant documents for the current order, and no content from other batches is included.
The values provided are common starting points and should be measured against the reader's own samples.