What the Data for This Category Looks Like
The primary stakeholders are electrode, assembly, formation, and quality departments of power, energy storage, and consumer battery enterprises. Daily documents include batching records, coating and rolling parameters, winding or lamination records, liquid injection records, formation and capacity testing data, OCV and safety test reports. Data sources include raw material batch records, MES systems, slurry and coating testing equipment, dew point monitoring devices, formation and capacity cabinets, and laboratory data. Data is updated in real time for each batch and individual cell. Each individual record contains specific parameter values and batch association information.
Constraints for Vector Models and Indexing
Lithium-ion cell manufacturing data is updated in real time per batch and individual cell. Indexes must support incremental updates to avoid resource consumption from full reconstruction. The data includes two categories: structured parameters and unstructured test reports. The system must balance vector recall accuracy for both types of content. Each data entry has strong parameter correlations. For example, coating and rolling parameters are directly linked to subsequent winding processes. Chunk splitting must preserve complete correlations between parameters and processes. Electrochemical systems are sensitive to parameters such as moisture content and cleanliness. Vector matching for related data must have high confidence to avoid recalling irrelevant records.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Single process record lengths for lithium-ion cell manufacturing documents mostly fall within this range, avoiding splitting that breaks links between parameters and processes |
maxToken | 8192 token | Adapts to the maximum supported token count of the bce-embedding-v1 model, matching vector generation requirements for chunk lengths |
defaultToken | 0.65–0.75 | Process parameter matching requires high confidence. This range filters out low-correlation non-target process records |
| Maximum Entries Per Search | Top 8 entries | Covers records for linked processes such as coating, winding, and formation, ensuring completeness of question and answer context |
| Maximum Context Window | 4000–6000 characters | Adapts to multi-process linkage requirements for process question and answer, avoiding context exceeding model processing limits |
| Reranking | Enabled | Sorts recall results by relevance, raising the display priority of key process parameters |
Three Common Misconfigurations
- Scenario: After uploading process data using the pushdata API in chunk mode, the FastGPT interface continuously displays the "Indexing" status with no progress updates. Cause: Incorrectly configured access permissions for
defaultToken, or chunk length exceeding themaxTokenlimit, resulting in failed indexing tasks. - Scenario: When calling the bce-embedding-v1 model, the returned error message is "This token is not authorized to use the model". Cause: The exclusive access key for this model was not bound in FastGPT, and an unauthorized key was used for the call.
- Scenario: When submitting a process specification question and answer, recall results lack parameter records for linked processes. Cause: Chunk length was set too large, causing key parameters to be split into different chunks and unable to be accurately recalled.
How to Verify Correct Configuration
- View the model management page of FastGPT v4.17.0, confirm that bce-embedding-v1 has been added and the
defaultTokenconfiguration is correct. - Upload a test coating and rolling parameter record, check the indexing task execution log to confirm there are no token over-limit or permission errors.
- Submit a process question and answer, such as "What is the range of areal density parameters for coating and rolling?", verify that the recall results include document fragments for the corresponding process.
- Adjust the similarity threshold, test recall results under different thresholds to confirm they meet the confidence requirements required by the business.
The values provided are common starting points and should be measured against the reader's own samples.