What data for this category looks like
Typical entities are cotton spinning, chemical fiber spinning, woven, and knitting enterprises. Daily documents include cotton blending lists, process setup sheets, equipment operation records, yarn inspection records, gray fabric inspection records, and defect records. Data sources cover raw fiber inspection reports, ERP systems, real-time spinning and loom data, evenness tester test results, and fabric inspection machine output data. Data update frequency shifts with production cycles. New documents generate or existing records update after each raw material batch, each production shift, and each bolt of fabric completion. Document fields include fiber type, setup tension parameters, yarn count specification, twist value, warp density, defect type, and more. Some fields use specialized units: twist is measured in twists per meter, and warp density is measured in ends per centimeter.
What constraints these characteristics impose on knowledge base retrieval and recall
Multi-source, heterogeneous data structures require the retrieval and recall link to adapt to both structured ERP parameter documents and unstructured inspection and record texts. This avoids limitations from a single retrieval mode. Frequently updated production data requires recall results to prioritize documents from the latest batches. This prevents use of outdated cotton blending or process parameters. Specialized fields require retrieval matching to focus on textile industry-specific terminology. This reduces irrelevant results from generalized recall. Business scenarios vary significantly across different documents. Retrieval needs differ between cotton blending lists and defect handling documents. Recall scope must be split by business module. This avoids invalid cross-scenario results.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the precision of textile-specific terminology, avoids recalling irrelevant defect records or equipment parameters |
limit | Top 8 entries | Covers multiple process documents tied to a single raw material batch, while keeping context length within FastGPT v4.17.0 context limits |
searchMode | mixedRecall | Adapts to structured ERP data and unstructured inspection records, balances parameter matching and text semantic understanding |
usingReRank | Enabled | Filters low-relevance defect descriptions from recall results, improves accuracy, and fits defect handling Q&A scenarios |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms for specialized terms like cotton blending and twist, improves recall coverage in professional scenarios |
collectionIds | Collection IDs divided by business module | Separates cotton blending documents, process setup documents, and inspection record collections, avoids invalid cross-module recall |
Three common configuration mistakes
- Scenario: Retrieval results are empty after enabling
usingReRank. Cause: Thesimilaritythreshold is set too high. This creates an insufficient initial candidate set, with no valid results after reranking. - Scenario: Retrieval results only include one document type, and cannot cover full-process process requirements. Cause: The
collectionIdsparameter is not configured. Recall scope is not split by business module, so only documents from the default collection are returned. - Scenario: Retrieval results include outdated process parameters. Cause: Sorting recall results by update time is not enabled. Latest batch production documents are not prioritized.
How to verify correct configuration
- Enter "Fiber matching requirements for cotton blending lists". Confirm recall results include cotton blending-related documents, and prioritize content updated in the last 7 days.
- After enabling
usingReRank, enter "Processing methods for unqualified yarn strength". Confirm irrelevant dyeing and finishing process documents are not present in recall results. - Adjust the
similaritythreshold. Verify recall results expand when the threshold lowers, and grow more precise when the threshold raises. - Review retrieval logs. Confirm
searchModeuses semantic retrieval + keyword hybrid mode, and matches both parameter keywords and text semantics.
The values provided are common starting points, and should be evaluated against the reader's own samples.