What the Data for This Category Looks Like
Core data for this category comes from several sources. These include orders and size charts, CAD patterns, raw and auxiliary material inspection records, hanging system data, needle detector test results, and customer sample sealing materials. Daily production outputs include process sheets, pattern files, layout plans, cutting lists, first-piece sample sealing records, process quality tables, and packaging size matching lists. Data updates follow a rhythm tied to each order, style, and batch. Individual documents contain fields such as process parameters, fabric color batch requirements, cutting piece matching rules, and size ratios. No unified fixed units exist. Specific units vary based on document type.
Constraints on Knowledge Base Retrieval and Recall
This category has high-frequency data updates split by order batch. Knowledge base retrieval must support incremental document synchronization. This avoids the time cost of full reconstruction. Multiple document types coexist in the dataset. The retrieval link must adapt to content in different formats. It must also set exclusive recall weights for core documents such as process sheets and layout plans. Individual documents have many exclusive fields and no unified standards. Retrieval must allow flexible matching of field keywords. It must also control redundancy in recall results. This prevents excessive irrelevant historical order data from interfering with current tasks. Cutting pieces have irreversible business characteristics. Recall results must have high confidence. This prevents incorrect process parameters from being introduced into current production links.
Configuration Settings
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires high matching accuracy for process parameters. It filters out irrelevant historical order data with low similarity |
limit | Top 8–12 entries | The total volume of production documents per batch is limited. Excessive recall will exceed the context window and reduce generation efficiency |
searchMode | mixedRecall | Balances structured field matching and keyword recall for unstructured process descriptions. It adapts to multiple document types |
usingReRank | Enabled | When multiple document types are mixed, re-ranking filters low-relevance results. It improves the priority of core process documents |
datasetSearchUsingExtensionQuery | Enabled | Expands the retrieval scope using current order keywords. It covers associated documents such as layout plans and cutting lists from the same batch |
collectionIds | Bind the current order's exclusive knowledge base collection | Isolates data by batch. It avoids cross-order document interference with the current production ledger summary task |
Three Common Mistakes
- Phenomenon: After uploading documents with images such as process sheets and layout plans, knowledge base retrieval results do not return content associated with the images. Cause: Relevant settings for image indexing are not configured, or the local deployment version does not enable image retrieval capabilities.
- Phenomenon: JSON parsing errors occur when calling the retrieval function over a network. Cause: The parameter format of the retrieval request does not meet specifications, or illegal characters in returned knowledge base results are not escaped.
- Phenomenon: Garbled text appears in retrieved and recalled text content when a thinking model generates production reports. Cause: The extra output switch of the thinking model is not turned off, or non-standard encoded characters in knowledge base documents are not preprocessed.
How to Confirm Correct Configuration
- Upload a single process sheet document from the current batch. Run a retrieval test, and check whether returned results include the core fields of this document.
- Check the binding relationship of the knowledge base collection. Confirm that the current retrieval task only associates the exclusive document collection of the current order.
- Adjust the value of the
similarityconfiguration item. Verify whether the number and relevance of retrieval results meet expectations. - Enable the
usingReRankfunction. Compare retrieval results before and after enabling, and confirm that the sorting priority of core process documents is improved.
The values provided here are common starting points and should be evaluated against the reader's own samples.