This configuration applies to FastGPT v4.17.0.
What This Category of Data Looks Like
Data for this category primarily comes from stamping workshops of automotive panel, home appliance sheet metal, and hardware parts enterprises. Daily documents include stamping work instructions, material utilization sheets, mold inspection sheets, first/last piece records, dimensional inspection reports, and mold change records. Data sources include steel coil labels and quality certificates, press and feeder operation data, mold management systems, MES systems, and coordinate measuring machine or fixture inspection data. Data is updated per coil, production batch, and mold change operation. Individual documents contain fields such as steel grade, material thickness, mold number, inspection dimensions, blank holder force parameters, with units including millimeters, kilonewtons, batch numbers, and others.
What Constraints These Characteristics Impose on Vector Models and Indexing
This category of data has high update frequency and scattered scenarios, including structured equipment operation parameters and unstructured quality inspection records and form documents. Data fields from different sources have large variations, so the vector model must be compatible with multi-format content to avoid field matching deviations during indexing. Incremental data from mold changes and production batches occurs frequently, so indexing construction must support incremental updates to avoid the time cost of full reconstruction. At the same time, data must be bound to key identifiers such as mold number and production batch, and the vector index must retain these associated fields to support precise associated retrieval in supplier and process quality audit scenarios.
How to Set the Configurations
| Configuration Item | Recommended Approach | Basis for This Approach |
|---|---|---|
| Chunk Length | 800–1200 characters | Most documents in this category are short-paragraph work instructions and inspection records. Overly long chunks will lose local contextual information. |
| Overlap Length | 100–150 characters | Documents such as mold change records and first/last piece records have continuous logic. Overlapping chunks retain contextual associations. |
defaultToken | 2048 | The average token count of single documents in this category matches this threshold, avoiding truncation of key parameters. |
maxToken | 8192 | Matches the maximum token limit of mainstream embedding models, avoiding model call errors. |
| Similarity Threshold | Calibrated based on actual testing | Matching accuracy requirements vary across different audit scenarios. Adjustments must be made based on actual retrieval results. |
| Single Search Limit | Top 8 results | Quality audits in this category require association with multiple related documents. Too many results increase subsequent processing time. |
| Maximum Context | 4000–6000 tokens | Audit scenarios require associating content from multiple documents such as mold, batch, and inspection data. Sufficient context supports complete reasoning. |
Three Common Misconfigurations
- The interface displays "No available channels" and a third-party embedding model API key is configured. The cause is that the knowledge base question and answer permission for the corresponding model was not enabled in group management, or the model's call channel was not configured correctly.
- Retrieval takes too long, and the knowledge base uses direct segmentation or automatic segmentation settings. The cause is that the chunk length is set too large or the overlap length is too high, resulting in excessive vector data volume for single-batch retrieval, or reasonable retrieval optimization strategies were not enabled.
- The interface displays "Model token limit exceeded" and a third-party embedding model is configured. The cause is that the set value of
maxTokenis less than the maximum supported token count of the model, resulting in a model call error.
How to Confirm the Configuration Is Correct
- Enter the vector model configuration page, check whether the added embedding model is shown as available, and verify that the API key and channel configuration match the actual deployment.
- Perform indexing construction for a small batch of documents, check whether field parsing errors or truncation prompts appear in the indexing progress log, and confirm the adaptability of the chunk parameters.
- Initiate a test query related to quality audit, check whether the associated fields of the retrieval results include key identifiers such as mold number and production batch, and confirm that the index association logic is normal.
- View the system's retrieval time consumption statistics, compare the time consumption changes before and after adjusting the configuration, and confirm that the parameter settings meet the scenario requirements.
The values provided here are common starting points and should be evaluated against the reader's own samples.