What the data for this category looks like
Typical entities in this category include garment OEM/ODM manufacturers, brand supply chain managers, and cutting and sewing workshops. Daily generated documents include process sheets, pattern instructions, layout plans, cutting orders, first sample sealing records, process quality sheets, and packaging and size matching sheets. Data sources cover orders and size charts, CAD pattern files, fabric and accessory inspection reports, hanging system process data, needle detector test records, and customer sample sealing requirements. The data update frequency is per order, per style, per batch. Each single document contains clear fields such as size, piece count, fabric batch, and process requirements.
What constraints these characteristics impose on vector models and indexes
This category has high data update frequency and strong correlation between single-batch documents, so vector indexes need to support incremental updates to accommodate frequent import of new orders and new styles. Different document types have large differences in text density: process sheets are structured text, while layout plans come with graphic metadata descriptions, so chunking logic for multiple document types must be supported. Retrieval needs focus on strongly related fields such as piece matching, size matching, and process verification, so the vector model must retain contextual associations between fields to avoid losing key matching logic after chunking. At workshop-level deployment, retrieval resource consumption must be controlled to avoid performance bottlenecks in concurrent scenarios.
How to set the configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapt to the text density of process sheets and layout plans, avoid including irrelevant content across processes in a single chunk |
maxToken | 4096 token | Match the input limit of general embedding models, compatible with long text chunks of garment-related documents |
| Overlap Length | 150–200 characters | Retain context for piece matching and size association, avoid losing matching logic across chunk retrieval |
| Similarity Threshold | 0.75–0.85 | Calibrated based on actual measurements, filter low-match non-target process requirements |
| Single Search Limit | Top 10 results | Control retrieval resource consumption, adapt to concurrent requirements at workshop level |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
Three common mistakes
- Phenomenon: A vector model connection failure prompt appears after local deployment, and the log shows a timeout. Cause: The
defaultTokenof the vector model was not correctly filled in the FastGPT configuration, or the oneAPI port mapping was not synchronized to the container. - Phenomenon: After importing the knowledge base, the retrieval results do not match the custom filter conditions, and a large number of irrelevant documents appear. Cause: The metadata of the order size chart was not associated in the index configuration, and vectors were only generated for plain text content, losing core filter fields.
- Phenomenon: After uploading a document, the system prompts an abnormal capacity calculation and cannot estimate storage requirements. Cause: Chunking parameters were not set separately by document type, causing the single chunk character count to exceed the preset range, leading to statistical logic deviations.
How to confirm the configuration is correct
- Upload a single complete process sheet, check that the vector generation progress bar in FastGPT v4.17.0 completes normally without error prompts.
- Enter the retrieval term "left front panel piece size deviation", verify that the returned document types are cutting orders or process quality sheets, and the match degree meets the preset threshold.
- Add a test order document, trigger incremental indexing, wait 1 minute, then retrieve the relevant content of the order, confirm that the results return normally.
- View the knowledge base capacity statistics panel, confirm that the total capacity matches the total character count of the uploaded documents, with no abnormal overflow.
The values given are common starting points and should be measured against the reader's own samples.