What this category of data looks like
Daily documents for this category include injection molding process cards, raw material drying records, mold history logs, first-piece inspection reports, patrol inspection forms, and material change and machine cleaning records. Data sources include resin batch and property sheets, injection molding machine parameters, mold temperature controllers, MES systems, and dimensional and appearance inspection equipment. Documents are updated per production batch, each mold or material change, and at the end of each shift. Document fields cover drying temperature, holding pressure, mold temperature, batch number, recycled material ratio, and more. Units include degrees Celsius, megapascals, kilograms, and others. The length of parameter sections in individual documents varies significantly.
What constraints these characteristics impose on vector models and indexing
Indexes must support incremental refresh to match the rhythm of mold or material changes and batch production, as data updates occur frequently with clear trigger points, avoiding resource consumption from full reconstruction. Document fields contain many numerical process parameters, so vector models must adapt to mixed text and numerical content to avoid splitting parameter associations. Large differences in the length of parameter sections in individual documents require retaining context of adjacent process steps during chunking, otherwise critical information such as packing and cooling windows and raw material drying conditions will be broken. Significant differences in field structures across different documents mean indexes must support filtering recall results by field dimensions to accurately match injection defect troubleshooting needs.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the typical length of process parameter sections in injection molding documents, avoiding splitting that disrupts associations between packing, cooling and other process steps |
| Overlap Length | 100–150 characters | Retains context across adjacent chunks, ensuring that parameter associations between packing windows and raw material drying records are not truncated |
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 | Adapts to the precision requirements of injection molding process parameters, filtering low-match irrelevant documents and avoiding deviations in defect troubleshooting |
| Single Search Limit | Top 10 entries | Covers the multi-dimensional process records required for injection defect troubleshooting, avoiding recall of excessive redundant data |
defaultToken | Calibrated via actual testing | Adapts to the token limit of the selected vector model, matching the total length requirements of imported documents |
Three Common Misconfigurations
- Phenomenon: When documents are imported, the generated dataset.csv only contains the index field and no content field. Cause: The document content extraction switch was not enabled, only index metadata was synchronized, and the process parameter content of injection molding documents was not fully captured.
- Phenomenon: When deploying FastGPT on an internal network, vector model calls experience timeout errors. Cause: No internal network model access proxy was configured, or timeout parameters were not adjusted to match internal network latency.
- Phenomenon: When importing long-section injection molding process cards using the default chunk length, parameter truncation occurs. Cause: Chunk length was not adjusted to fit the long parameter sections of injection molding documents, leading to critical information such as packing and cooling windows being split.
How to Confirm Proper Configuration
- Import a single injection molding process card, view the chunk preview interface of the knowledge base, and confirm that process parameters are not truncated and associated information of adjacent paragraphs is fully retained.
- Export the dataset.csv file of the current knowledge base, check that the file contains both index and content fields, and that the content field contains complete process parameter content.
- Initiate a query for injection defects, confirm that the recall results include content from relevant documents such as corresponding mold history logs and raw material drying records.
- View the FastGPT system logs, confirm that there are no timeout errors in vector model calls, and that index refresh operations complete according to the preset rhythm.
The values provided are common starting points and should be evaluated against the reader's own samples.