What data for this category looks like
The data for this category mainly comes from concentrate and auxiliary material analysis reports, DCS systems, electrolytic cell voltage and current monitoring, laboratory test results, energy and environmental ledgers. Daily documents include batching lists, smelting records, electrolytic cell condition records, anode and cathode quality records, environmental ledgers, and similar items. Most fields include element content, process parameters, batch numbers, timestamps, and similar details. Update frequency changes in real time with raw material batches, furnace heats, or cell conditions. Some summary ledgers are updated on a fixed cycle.
Constraints imposed by these characteristics on vector models and indexing
The multi-source heterogeneous data structure of this category requires the index system to support mixed-format embedding and association. It must enable unified indexing of real-time numerical values, structured records, and text reports. The batch-based production data update rhythm requires the index to support incremental refresh by production batch, to avoid performance loss from full reconstruction. The abundant process parameters and element detection fields require the vector model to adapt to embedding of multi-dimensional numerical and text features. It must also retain associated fields such as batch numbers and cell numbers, to ensure that full-link data for the same batch can be accurately associated during traceability.
Configuration guidelines for FastGPT v4.17.0
Refer to the following configuration reference for FastGPT v4.17.0:
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-page text length of smelting records and laboratory reports, avoids information fragmentation within chunks |
| Overlap Length | 150–200 characters | Retains cross-chunk process parameter associations, ensures continuity of cell condition and batch information |
| Similarity Threshold | 0.72–0.85 | Calibrated via actual testing, adapts to similarity matching for numerical features such as element content and cell voltage |
defaultToken | 8192 | Adapts to the total token length after splicing multiple batches of data, meets context requirements for full-link traceability |
maxToken | 16384 | Reserves expansion space, supports multi-cell data association queries for complex electrolytic cell groups |
| Single Search Limit | Top 10 entries | Balances recall accuracy and query efficiency, covers the main associated data for a single batch |
Three common misconfigurations
- Symptom: Vector model calls return a 401 error, curl tests work normally. Cause: The vector model API key or access whitelist is not configured correctly, and FastGPT's request headers do not carry valid authentication credentials.
- Symptom: Duplicate batch records remain after index merging. Cause: The deduplication switch is not enabled in the index configuration, or the unique identifier field used for deduplication (such as batch number, cell number) is not specified.
- Symptom: Domestic vector models cannot be loaded. Cause: The access address of the corresponding model is not added in FastGPT, or the model's input format does not match FastGPT's embedding requirements.
How to confirm successful configuration
- Upload a single smelting record and laboratory report, verify that the number of index-generated chunks matches the configured chunk length.
- Submit a query that includes cell number and batch number, confirm that the returned results include the associated fields specified in the configuration.
- After connecting the vector model, use the built-in test tool to generate embedding vectors, compare whether the results are consistent with external call returns.
- Perform an index merge operation, check whether the number of duplicate records in the result matches the configured deduplication rules.
The values listed above are common starting points, and should be validated against local samples.