What the Data for This Category Looks Like
Technical documents for complete home appliance manufacturing include bill of materials (BOM), assembly work instructions, refrigerant charging records, electrical safety test reports, random performance inspection reports, and repair records. Data is sourced from ERP/MES systems, component batch ledgers, tightening and charging device logs, safety compliance testers, performance laboratory data, and after-sales fault archives. Updates run per single product, per shift production batch, or per random inspection plan. Each document includes fields such as model specification, process parameters, test results, and batch number. Some documents include physical units such as megapascals, degrees Celsius, and milliamps.
What Constraints These Characteristics Impose on Vector Models and Indexing
A single home appliance corresponds to multiple associated documents. Data updates occur frequently, and batch association is strong. This requires indexes to support fast, accurate recall by batch and model. Documents contain large amounts of structured process parameters and unstructured test descriptions. Vector models must adapt to encoding needs for mixed content. Compliance requirements for safety and performance test documents mean vector recall accuracy must match process tolerance ranges, to avoid risks of wrong assembly due to mixed production lines. In mixed-line production scenarios with multiple models, indexes must support dimension isolation by product model, to prevent cross-interference between documents of different models. High-frequency updates for batch documents also require index incremental synchronization efficiency to match shift production rhythms.
How to Configure the Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph text length of assembly work instructions and test reports, avoids splitting that breaks the correlation of process parameters |
maxToken | 4096 token | Matches the encoding requirements of long sections of performance test data in home appliance documents, avoids truncating key parameters |
| Similarity Threshold | 0.75–0.85 | Distinguishes differences between documents of the same model with different batches, prevents incorrect model matching during mixed-line assembly |
| Single Search Limit | Top 10 results | Covers the number of associated documents for a single product, ensures comprehensive recall |
defaultToken | 1024 token | Adapts to the encoding length of short documents such as component batch records, balances calculation efficiency and recall accuracy |
| Reranking | Enabled | Optimizes the ranking of recall results, prioritizes core review documents such as safety compliance and performance test reports |
Three Common Misconfigurations
- Phenomenon: Only one NVIDIA 3090 GPU is detected in the Docker container. Video memory usage during vector model loading equals only the capacity of a single card, and the second card has no load. Cause: The
--gpus allparameter is not added to the docker run command, or multi-GPU parallel scheduling is not enabled in the FastGPT system configuration. - Phenomenon: A 404 Not Found error is returned when calling the Baidu embedding-v1 interface. Logs show the interface path does not exist. Cause: The model mapping path of OneAPI is not configured correctly, or call permission for the corresponding model is not activated.
- Phenomenon: Attempting to add a vector model fails in a non-GPU environment, and the system prompts that the hardware does not support CUDA acceleration. Cause: A CPU-adapted vector model is not selected, or the CPU running mode is not specified in the container startup command.
How to Confirm the Configuration Is Correct
- Upload a typical home appliance assembly work instruction, check the vector index splitting results, and confirm that no key process parameters are truncated.
- Initiate a technical document search for a specific home appliance model, confirm that the recall results include the BOM, safety compliance report, and assembly work instruction for the corresponding model, and the number of results matches expectations.
- View the FastGPT system hardware monitoring panel, confirm that multiple GPUs are being called normally, and video memory usage matches configured expectations.
- Test calling the configured vector model interface, confirm that the returned vector data is normal, and no error logs are generated.
The values provided here are common starting points, and should be evaluated against the reader's own samples.