What the data for this category looks like
This category of data is sourced 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. Data update cycles cover full-process records for each production batch and individual cell, as well as real-time operating data. Daily documents include batching records, coating and rolling parameter ledgers, winding/lamination operation records, liquid injection process logs, formation and grading operation data, OCV and safety test reports. Data fields include slurry ratio parameters, coating areal density, winding tension, liquid injection volume, internal resistance values, formation and grading curves, etc., with corresponding units such as g/㎡, mΩ, Ah, and others.
What constraints these characteristics impose on the "multi-turn dialogue and prompt words" link
The full-process data of lithium-ion cell manufacturing has multiple dimensions and is closely linked. Single-batch defect traceability requires associating cross-document records from multiple links, which requires multi-turn dialogue to continuously call historical records from different processes. The update cycles of real-time operating data and batch archived data differ significantly, so prompt words need to distinguish the calling logic between real-time queries and historical traceability. Electrochemical systems are sensitive to moisture and cleanliness, so defect analysis requires accurate matching of batch parameters, and prompt words need to clearly define field ranges to filter irrelevant information. The full-process data volume of a single cell accumulates with production links, so multi-turn dialogue needs to reasonably limit the context length to prevent exceeding processing limits.
How to set the configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Number of chat records | First 6 entries | Matches the core process links of a single cell batch, avoids redundant data interfering with analysis |
| Maximum context token count | 8000–12000 token | Adapts to the accumulated length of full-process cell data, prevents truncation of critical parameters |
chatId | Bound to production batch ID | Accurately locks the full-link data of the corresponding batch to enable cross-process traceability |
messages | Spliced by process group | Organizes historical conversations in the order of coating, winding, and formation, which aligns with business analysis logic |
variables | Bound to cell batch number, inspection items | Limits the query scope to avoid calling irrelevant workshop-wide data |
Three common mistakes
- Embedding a text content extraction node in the workflow and configuring 6 chat records fails to trigger multi-turn dialogue. The cause is that the workflow node did not correctly bind the
chatIdparameter, making it impossible to associate historical dialogue context. - Format errors occur when calling a multimodal model to process cell inspection images. The cause is that the prompt words did not clearly specify the input format requirements for images, causing the model to fail to correctly parse the incoming image data.
- The online version of the dialogue remains in a state of no response when searching the knowledge base. The cause is that the multi-turn dialogue context does not limit a reasonable length, causing the system to load excessive redundant data and trigger a timeout.
How to confirm the configuration is correct
- Enter the production batch number to initiate a query, and check whether the returned results only associate the corresponding process data of this batch.
- Initiate two consecutive queries for different processes, and check whether the system can associate the previous dialogue context for subsequent analysis.
- View the parameter configuration of the workflow node, and confirm that the
chatIdandmessagesparameters have been correctly bound to business fields. - Trigger a multimodal image query, and check whether the system can normally receive and parse the uploaded inspection image files.
The values provided are common starting points and should be measured against the reader's own samples.