What data for this category looks like
Data for this category comes from raw material batch ledgers, production MES systems, slurry and coating testing equipment, workshop dew point monitoring devices, formation and grading cabinets, and laboratory testing equipment. Daily generated documents include batching records, coating and rolling parameters, winding or lamination records, liquid injection records, formation and grading data, OCV and safety test reports. Data update rhythm falls into three categories: batch documents such as batching and liquid injection are updated after each production batch is completed, single-cell documents such as formation and grading data and OCV are updated after each individual cell test is completed, and workshop environmental parameters are updated in real time. Document fields include raw material batch number, feeding weight, roll gap value, linear speed, liquid injection volume, charging current, open circuit voltage, dew point temperature, and more. Corresponding units include kilograms, millimeters, meters per minute, milliliters, amperes, millivolts, and degrees Celsius.
What constraints these characteristics impose on the multi-turn dialogue and prompt engineering workflow
Data for this category is updated in three categories: batch, single-cell, and real-time, and covers multiple types of production and test documents, with corresponding fields containing core parameters from different production links. This requires that multi-turn dialogue must first guide the clarification of the review target scope, such as specifying a cell batch number or single-cell ID, to avoid cross-dimensional data confusion. At the same time, extraction rules for each field must be clearly defined in the prompt, distinguishing the field definitions and units of different documents to prevent parameter extraction deviations. Real-time updated workshop environmental parameters must be included in the verification logic during dialogue to ensure the timeliness of cross-review. In addition, multi-turn dialogue must support the gradual breakdown of complex review requirements, such as first confirming slurry batching records, then verifying the matching of coating and rolling parameters.
How to configure settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Context window length | 8000–12000 characters | Covers multiple production and test documents for a single lithium-ion cell batch, avoiding truncation of core parameter content |
variables parameter configuration | Bind batch numbers, cell IDs, and core parameter fields | Data for this category is centered around production batches and single cells, so clear variable anchors are required to accurately extract review targets |
messages historical context processing | Retain the most recent 5 rounds of dialogue content | Only recent logic and parameters are needed during multi-turn review interactions, avoiding redundant data interfering with review judgments |
chatId session isolation configuration | Assign independent sessions by production batch | Production data for different batches is independent, so cross-batch session data confusion must be avoided |
| Variable extraction rule configuration | Define by document type module, specifying fields and corresponding units | Documents for this category include multiple types of production parameters, with differing fields and units, so clear extraction rules are required to avoid deviations |
| Multi-turn interaction timeout setting | 600 seconds | Review requirements in production scenarios typically do not require long wait times; timeout prevents session resource occupation |
Three common mistakes
- Symptom: Target parameters cannot be accurately extracted during multi-turn dialogue, such as feeding weight in batching records or roll gap values in coating and rolling operations. Cause: No clear extraction rules are defined for the category-specific document fields and units, leading the model to confuse the meanings of different parameters.
- Symptom: Cross-batch production data appears in sessions, leading to incorrect cross-review results. Cause: Sessions are not isolated by production batch via the
chatIdconfiguration, so context data from different batches is incorrectly associated. - Symptom: Production parameters entered via voice input cannot be correctly converted to text, or the converted text lacks key unit information. Cause: No specialized terminology and unit rules for the lithium-ion cell manufacturing field are included in the prompt, so the speech-to-text model cannot accurately recognize exclusive parameters.
How to confirm configurations are properly set up
- Initiate a test dialogue, enter a known production batch number and corresponding document content, verify that the variable extraction results include correct fields and values, and adjust extraction rules until expectations are met.
- Initiate two test sessions for different batches, isolate them via the
chatIdconfiguration, verify that review data from the two sessions does not mix, confirming that session isolation is active. - Initiate a multi-turn interaction test, verify that the context window only retains recent interaction logic and parameters, with no redundant historical data interfering.
- Test the voice input function, enter specialized production parameters, verify that the transcribed text can be correctly identified as corresponding fields, and adjust prompt rules until requirements are met.
The values provided are common starting points and should be measured against the reader's own samples.