What the Data for This Category Looks Like
Core data for this category comes from customer product models, mold flow or forming simulation files, CAD/CAM exported drawings, machining and measurement records, and prototype part data. Daily output documents include mold design task sheets, DFM reports, part drawings, machining process cards, mold trial reports, mold modification notices, and mold history records. Data updates proceed on a per-project basis. New version files generate after each mold trial or modification. Data fields include mold number, part dimensional tolerances, machining parameters, mold trial defect records, number of mold modifications, and more. Some files contain both structured parameters and unstructured text content.
Constraints Imposed on Vector Models and Indexing
This category’s data updates iteratively per project and includes both structured parameters and unstructured design content. Indexes must support associating multi-version data by project dimension. They must adapt to incremental synchronization after a single mold trial or modification. Structured fields such as part dimensional tolerances and machining parameters require association with non-text content for retrieval. Indexes must support combined matching of metadata and vector text. Mold data from different projects has weak correlation. This requires avoiding irrelevant recall results across projects. Index recall ranges must support filtering by project dimension. Vector conversion of large volumes of engineering drawings and simulation files must adapt to unstructured text segmentation rules. This prevents loss of key design details.
Configuration Settings
This guide applies to FastGPT v4.17.0. Below are specific configuration recommendations:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to continuous text content such as part dimensions and machining processes in mold design documents, avoiding splitting key design paragraphs |
| Overlap Length | 50–100 characters | Retains associated details of adjacent chunks, preventing truncation of mold trial defect descriptions and mold modification-related information |
defaultToken | 4096 | Matches the single-page text length limit of most mold design documents, and adapts to vector model input constraints |
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 | 0.72–0.85 calibrated via actual testing | Filters low-match irrelevant mold project data, avoiding cross-project recall interference with audit processes |
| Single Search Limit | Top 8 entries | Controls the number of recall results, adapting to mold audit scenarios where retrieval must focus on core issues |
Three Common Misconfigurations
- Symptom: The dataset.csv generated from imported mold design documents only includes the index field, with no content field data. Cause: Incorrect configuration of chunk length and text extraction rules results in unextracted or unwritten text content from unstructured documents to the target field.
- Symptom: After integrating a vector model in a Docker deployment environment, retrieval results are empty or timeout errors occur. Cause: Incorrect configuration of model access ports and network policies prevents the FastGPT container from connecting to the vector model service, or the
defaultTokensetting exceeds the model input limit, causing a timeout. - Symptom: Retrieval results include irrelevant mold data from cross-projects, unable to accurately match the current audit project. Cause: Project dimension-based index filtering configuration is not enabled, so the recall range covers all historical project data and interferes with the current audit process.
How to Confirm Proper Configuration
- Review knowledge base import logs to confirm that all mold design document content fields have been correctly extracted and written to the vector database.
- Submit a retrieval request containing mold numbers and mold trial defect keywords, and verify that recall results only include associated documents from the current audit project.
- Adjust chunk length and overlap length parameters, and verify that vector conversion results under different configurations retain key design details.
- Check vector model access logs to confirm that the FastGPT container can normally connect to the model service, with no timeout or error records.
The values provided here are common starting points, and should be evaluated against in-house sample data.