What data looks like for this category
Data for this category primarily comes from garment OEM/ODM workshops and brand supply chain links. Daily documents include process sheets, CAD patterns, layout plans, cutting orders, first sample seals, process quality forms, and packaging size matching sheets. Data sources cover order and size charts, fabric and accessory inspection records, hanging system operation data, needle detector test results, and customer sample seal documents. Data updates are triggered with each order, each garment style, and each production batch. Document fields include specific sizes, fabric color numbers, cutting piece numbers, process requirements, and more. There is no unified fixed field order, and configurations are dynamically adjusted according to order requirements.
What constraints these characteristics impose on vector models and indexes
Data for this category updates frequently with orders, styles, and production batches. This requires indexes to support incremental synchronization and fast updates. Document fields are scattered and lack unified formats. Some documents contain coordinate parameters and cutting ratio data from CAD patterns, as well as process node data from hanging systems. Vector models must adapt to mixed content of non-standard text and structured parameters. The irreversible nature of cutting pieces requires retrieval results to accurately match process requirements and size specifications, avoiding production deviations caused by fuzzy matching. Format differences across multiple data sources require index configurations to be compatible with different types of text and numerical inputs.
How to set configurations
These configuration rules are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapt to the single-section content length of documents such as process sheets and layout plans, avoid splitting cross-process information |
| Overlap Length | 100–200 characters | Retain association information between processes, prevent cutting piece matching requirements from being truncated during chunking |
| Similarity Threshold | 0.75–0.85 | Meet the requirement of accurately matching precise fields such as garment sizes and color numbers, filter low-relevance maintenance reference content |
maxToken | 4096 token | Adapt to long text content such as equipment inspection manuals and long process descriptions, ensure complete loading of maintenance steps |
| Single Search Limit | Top 10 results | Cover maintenance references for different sizes of the same style, avoid difficulty in filtering caused by too many results |
| Reranking | Descending by retrieval score | Prioritize content with the highest matching degree to equipment models and process requirements |
Three common mistakes
- Retrieval result sorting does not meet expectations. Some content with high matching degree to the current order size and process is ranked lower. Cause: Secondary reranking for process and size fields is not configured, only initial retrieval scores are relied on for sorting, which does not adapt to the precise matching needs of the garment category.
- "undefined model" error occurs when calling the vector model. Cause: An officially supported vector model is not selected, or the input model name format does not meet requirements. Some non-standard models cannot be recognized by the system.
- Auxiliary data is incorrectly included in the index, leading to retrieval redundancy. Cause: Core maintenance documents and auxiliary fabric and accessory inspection records are not distinguished. Non-maintenance-related auxiliary data is vectorized together, increasing invalid retrieval results.
How to verify configurations are properly set
- Upload a process sheet that includes specific sizes and process requirements. Check whether the chunked text in the system retains key information such as cutting piece matching and color number requirements intact, to verify the effect of chunk length and overlap length configurations.
- Enter a query related to equipment inspection. Check whether the sorting of retrieval results prioritizes matching the current order style and size, to verify the rationality of the similarity threshold and reranking configurations.
- Submit new maintenance documents. Check whether the index completes updates within a reasonable time, to verify the effectiveness of incremental synchronization configurations.
- Test calls of different officially supported vector models, confirm that there are no model format error prompts, to verify the correctness of model configurations.
The values listed here are common starting points, and should be evaluated against samples from the target deployment scenario.