What the Data for This Category Looks Like
The data for this category originates from raw material batch ledgers, workshop MES systems, slurry and coating testing equipment, dew point monitoring devices, formation and grading cabinets, and laboratory testing equipment. Daily documentation includes batching records, coating and rolling parameters, winding or stacking records, liquid injection records, formation and grading data, OCV and safety test reports. Fields cover process parameters, test values, and batch identifiers, with units aligned to process specifications for each stage. For example, coating areal density uses g/㎡, and dew point uses ℃. Data updates occur per batch, per individual cell, or in real time. The volume of documents per single batch grows with cell production scale.
What Constraints These Characteristics Impose on the Vector Model and Indexing Stage
Lithium-ion cell manufacturing data is associated by batch and updated frequently. Indexes must support incremental updates and fast recall at the batch level. Documents include both structured parameters and unstructured test reports. A balance must be struck between parameter integrity and text coherence during chunking. Electrochemical systems have extremely high requirements for parameter consistency. Retrieval must prioritize matching associated data from the same batch and same process, to avoid low-correlation results interfering with anomaly judgment. Long-text formation and grading curve reports must also adapt to the length limits of vector models, to prevent key data from being truncated.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale for This Selection |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of single-batch process records, avoiding splitting critical parameter groups |
| Overlap Length | 100–150 characters | Retains cross-chunk process parameter associations, preventing chunking from breaking linked coating and rolling data |
| Similarity Threshold | 0.75–0.85 | Matches the strict requirements for parameter consistency in cell manufacturing, filtering low-correlation abnormal reports |
| Single Search Limit | Top 10 results | Limits the recall scope, focusing on relevant documents from the same batch or same process |
maxToken | 4096 token | Adapts to long-text formation and grading curve reports, avoiding truncation of critical data |
defaultToken | Calibrated based on actual measurements | Adapts to document lengths across different stages, covering full-process data from batching to testing |
Three Common Configuration Errors
- Issue: An "incomplete parameters" error is returned when creating a training order. Cause: The cell batch number was not passed as required metadata to the training interface, making it impossible to associate documents by batch.
- Issue: Retrieval results after vectorizing images cannot match corresponding process parameters. Cause: The bound index between image OCR and structured text was not enabled; only image text was extracted without associating process fields.
- Issue: No options appear when selecting a target Embedding model. Cause: The model list was not updated in the oneAPI configuration, and the interface key and address for the corresponding model were not added.
How to Verify the Configuration Is Properly Set Up
- Upload a single batch's batching record and abnormal report, check whether batch metadata is automatically associated after the training order is generated.
- Upload an image containing process parameters, verify during retrieval that both the image OCR text and corresponding structured parameter documents are returned.
- On the model configuration page, confirm that the target Embedding model appears in the optional list, and that interface connectivity is normal.
- Initiate an abnormal report retrieval, check that the similarity threshold of recall results falls within the preset range, with no low-correlation results.
The values provided are common starting points and should be evaluated against the reader’s own samples.