Data for This Category
Daily data for this category includes process sheets, CAD patterns, fabric layout plans, cutting orders, first article sealed samples, process quality sheets, packaging size matching sheets, and similar items. Data sources primarily include orders and size charts, CAD design files, fabric and accessory inspection records, hanging production systems, needle detector machine data, and customer sealed samples. Data updates follow the rhythm of individual orders, individual styles, and individual batches. This category has a large number of styles and strong seasonality. Fabric cut pieces are irreversible, and manual processes are labor-intensive. No universal standardized batch of physical data exists for this category.
Constraints for Vector Models and Indexes
This category has diverse data types, including structured tables, CAD drawing files, text records, and device-collected data. This requires vector models to support non-standardized multi-source inputs. Data updates occur frequently, and order variations per style are significant. Indexes must support fast incremental updates to align with each production batch's rhythm. The irreversible nature of fabric cut pieces means vector recall accuracy must cover process details and pattern matching to avoid batch deviations. Additionally, the labor-intensive manual process feature results in fine field granularity for text data such as process quality sheets. Chunking must align with process splitting logic to avoid disrupting process-related information.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-process description length of process quality sheets and process sheets, avoids splitting that disrupts process logic |
| Overlap Length | 100–150 characters | Covers associated descriptions of adjacent processes, prevents key process information from being truncated during chunking |
defaultToken | 8192 | Adapts to the text annotation length of CAD patterns and fabric layout plans, meets context accommodation needs for multi-segment process data |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
| Similarity Threshold | 0.75–0.85 | Filters low-match non-target process data, avoids false recalls of batch errors such as size and color difference issues |
| Single Search Limit | Top 10 entries | Matches the number of processes for a single-style order, avoids redundant results interfering with on-site audit judgments |
Three Common Misconfigurations
- Symptom: Index creation fails. The host machine is an 8c16G non-GPU virtual machine with small file sizes. Cause: Chunking parameters were not adjusted for multi-source unstructured data, leading to post-chunking tokens exceeding the model's carrying limit and triggering an out-of-memory error.
- Symptom: An error occurs when calling the embedding model, and the specified channel cannot be accessed. Cause: The index model's calling channel was not configured separately, and the global chat model configuration was used by default, resulting in cross-channel model compatibility issues.
- Symptom: A 503 error is returned after starting the container, prompting that there are no available channels for text-embedding-v3 under the default group. Cause: The API key and channel configuration for the index model were not correctly injected in the docker-compose configuration, resulting in an interruption of the model call link.
How to Verify Proper Configuration
- Upload the process sheet and process quality sheet for a single clothing style, check that chunked text fragments retain complete process descriptions with no key information truncated.
- Initiate a pattern matching retrieval, verify that the similarity interval of returned results falls within the preset threshold range, and no irrelevant historical order data is included.
- Start an incremental index update, verify that newly uploaded batch order data can be quickly added to the existing index, with no long wait times or failure prompts.
- View the FastGPT v4.17.0 container logs, confirm that there are no channel errors in embedding model calls, and that the API key configuration matches the index model.
The values provided are common starting points and should be evaluated against the reader's own samples.