What the data for this category looks like
Data for this heat treatment scenario comes from daily business documents and device-collected data of professional heat treatment plants. This includes heat treatment process cards, furnace loading diagrams, furnace temperature curves, hardness and metallographic reports, equipment inspection sheets, exception handling forms, and more. Data sources cover material grades, customer technical requirements, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers and metallographic equipment, and manufacturing execution systems (MES).
The data update frequency is per furnace load and per batch. A single data entry contains structured parameters and unstructured text. Fields include temperature, time, hardness value, metallographic structure grade, and more. Units include degrees Celsius, hours, Vickers hardness (HV), structure grade, and more.
What constraints these characteristics impose on the "vector models and indexes" link
This category of data includes multiple types: time-series data (such as furnace temperature curves), structured parameters (such as process card temperature and holding time), report-style text, and others. The field span of a single data entry is large. When chunking, context association of key business fields must be retained to avoid losing the corresponding relationship between process parameters and test results after splitting.
The data update frequency is per furnace load and per batch. The incremental index trigger logic must adapt to high-frequency small-batch update scenarios to avoid redundant calculations from full indexes. The data includes numeric fields such as temperature and hardness values. The vector model must support mixed input of numeric features and text features, otherwise vector representation deviation will occur.
The bidding scenario requires associating customer technical requirements with internal process capabilities. The index must support cross-document associated recall to ensure retrieval results cover multi-dimensional business information.
How to set the configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the text length of heat treatment process cards and furnace temperature curves, avoiding loss of association between process parameters and test results after splitting |
| Overlap Length | 150–200 characters | Retains temporal or parameter context between adjacent chunks, preventing segmentation breaks in furnace temperature curves |
| Similarity Threshold | 0.72–0.85 | Matches text similarity between customer technical requirements and internal process parameters, filtering low-match invalid data |
| Single Search Limit | Top 8 entries | Balances recall coverage and calculation efficiency, adapting to multi-document association needs during bidding |
defaultToken | 4096 | Adapts to the input length of locally deployed embedding models, avoiding information loss from text truncation |
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. |
Three common errors
- A locally Docker-deployed M3E model cannot load properly, and the interface prompts that the model was not found. The cause is failure to map the local model port to the FastGPT container network, or the configured model access address does not include the correct port number.
- After repeatedly configuring an embedding model with the same name, the original configuration is overwritten, and index recall results show abnormal fluctuations. The cause is that FastGPT's model configuration rules only allow a unique model name to correspond to one set of parameters, and different deployment instances are not assigned differentiated names.
- A large number of empty fields appear in chunked Excel source data (such as equipment inspection sheets), and key parameters are missing from retrieval results. The cause is that the chunk length is set too large, which does not adapt to the compact structure of single-row Excel data, resulting in truncated association between headers and corresponding rows during splitting.
How to confirm successful configuration
- Enter the FastGPT model management page, check the configuration status of the local embedding model, and confirm that it displays "Connected".
- Upload a single heat treatment process card as test source data, perform chunking and index construction, and verify that the chunk list retains key fields such as heating temperature and holding time.
- Input a test query such as "What process scheme meets a customer's requirement for quenching hardness of HV500", and check whether the number and matching degree of retrieval results match the preset threshold.
- Try adding a new furnace temperature curve data to trigger incremental indexing, and confirm that there are no error messages in the index update log.
The values provided here are common starting points, and should be evaluated against your own samples.