This document is written for FastGPT v4.17.0.
What the data for this category looks like
The data for this category mainly comes from injection molding, stamping, and die-casting mold enterprises, as well as mold centers of manufacturing enterprises. Daily documents include mold design task sheets, DFM reports, part drawings, processing technology cards, mold trial reports, mold modification notices, mold history records, and other related documents. Most data is sourced from customer product models, mold flow or forming simulation files, CAD/CAM output files, machining and measurement data, and scan or measurement results of trial-molded samples. Updates follow individual project progress, with new data generated after each mold trial or mold modification. Most fields include geometric parameters, process requirements, defect records, iteration versions, and other relevant content.
What constraints these characteristics impose on the vector model and index link
This category has diverse data types, including structured parameters, unstructured text, and multimodal content. This requires vector models to support encoding adaptation for mixed data types. Data is updated frequently and iterated per project, so indexes must support incremental updates and version differentiation to avoid mixing documents from different projects. Fields vary significantly across different documents: for example, mold trial reports focus on defect records, while process cards focus on parameter requirements. Chunking logic must adapt to the integrity requirements of different content to avoid damaging key information. Mold inspection involves precision and compliance requirements, so vector recall must accurately match process parameters and defect records, which creates clear precision requirements for similarity threshold settings.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxToken | 8192 | Adapts to long text process cards and mold trial reports in mold documents, avoiding truncation of key parameters |
| Chunk Length | 800–1200 characters | Mold design documents include geometric descriptions and process requirements; this range retains complete process unit information |
| Overlap Length | 100–150 characters | Avoids breaking contextual associations of mold trial defect records across chunks |
| Similarity Threshold | 0.75–0.85 | Requires accurate matching of process parameters and defect records, avoiding recall of unrelated iterative version documents |
| Single Search Limit | Top 10 entries | Mold inspection needs to cover multi-dimensional process and defect information; appropriate recall ensures comprehensive review |
defaultToken | 2048 | Adapts to the average length of daily inspection documents, balancing encoding efficiency and content integrity |
Three Common Misconfigurations
- Phenomenon: The interface returns an error indicating failure to call the multimodal vector model, with a prompt that the model is not supported. Cause: The access parameters for the corresponding multimodal model were not added in the platform configuration, or the access rules of different manufacturers' vector models were mixed.
- Phenomenon: The number of index recall results does not match the set single search limit, with too many or too few recalled contents. Cause: The associated logic between the similarity threshold and single search limit was not configured correctly, or the chunk length setting was unreasonable leading to abnormal chunk counts.
- Phenomenon: Custom RAG queries return empty results, failing to match relevant mold documents. Cause: The index mapping was not configured for specific fields of mold documents, such as the version number of mold modification notices, or key process parameter information was truncated during vector encoding.
How to Confirm the Configuration Is Correct
- Upload a single mold trial report, perform vector encoding, and check the chunking results to confirm that key process descriptions are not truncated.
- Initiate a query containing specific process parameters, verify the document type and content matching degree of the recall results, and adjust the similarity threshold to a range that meets business requirements.
- Upload a new mold modification notice, trigger incremental indexing, and check the index update log to confirm that the task executed successfully.
- Test the configured vector model access, verify that there are no format errors during calls, and confirm that the model parameter configuration is correct.
The values provided are common starting points and should be measured against the reader's own samples.