What data for this category looks like
Data for this category comes from winding, insulation, final assembly, and testing departments of electric motor factories and power equipment enterprises. It covers winding process cards, insulation treatment records, dynamic balance reports, voltage withstand insulation test reports, and type and factory inspection test reports. Data sources include electromagnetic design data, copper wire and insulation material batch records, winding equipment data, balancing machine data, electrical test bench data, and ERP and MES system information. Data is updated after each product or batch production, and synchronized whenever product type specifications are changed. Documents include dedicated fields such as winding turns, insulation parameters, and test values, with associated identifiers such as batch numbers and equipment numbers.
What constraints do these characteristics impose on the multi-turn dialogue and prompt engineering link
Data for this category is split by batch, equipment, and test type. Multi-turn dialogue must associate identifiers such as batch numbers and equipment numbers to avoid mixing winding and insulation data from different batches. Documents include dedicated professional fields related to electromagnetic, insulation, and mechanical balance content. Prompt engineering must clearly limit the retrieval scope to technical documentation for electric motor and generator manufacturing to prevent mixing in data from unrelated categories. Data is updated frequently, so multi-turn dialogue context must filter expired test records and batch data. Additionally, documents involve sensitive content such as high-voltage tests and temperature rise tests. Prompt engineering must restrict the LLM to generate responses only based on retrieved document content, and prohibit introducing external unvalidated information.
Configuration Settings
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Single search limit | Top 5 entries | Technical documents for this category are split by batch and test type. The number of relevant documents to retrieve per dialogue turn is limited. The first 5 entries can cover most associated data, avoiding redundant information interfering with the cross-review process. |
| Question optimization | Enabled, calibrated based on actual testing | This category includes professional terms such as electromagnetic, insulation, and mechanical balance. Question optimization can convert user professional queries into expressions better matched for knowledge base retrieval, improving recall accuracy. |
| Maximum context | 8000–12000 token | Documents for this category include multiple sets of test data and batch information. Multi-turn dialogue context must retain sufficient historical dialogue and retrieval results to avoid truncation of critical review information. |
variables | Pass batch numbers, equipment numbers | Data for this category is associated via batch and equipment identifiers. Passing dedicated identifiers via variables allows the LLM to accurately locate corresponding documents, avoiding mixing winding and insulation data across batches. |
messages | Retain the most recent 3 dialogue turns | During multi-turn review, only the most recent interactions and retrieval results need to be retained in historical dialogue. Excessively long context increases LLM processing load, while also avoiding expired information interfering with current review. |
chatId | Bind to the current dialogue session | During multi-turn review, chatId can associate all interaction records for the current session, ensuring context consistency and avoiding mixing review data from different sessions. |
Three Common Mistakes
- Phenomenon: Retrieved document content includes paragraph numbering or id fields, which interfere with cross-review judgments. Cause: Document metadata fields are not filtered via configuration items, causing unrelated identifier information to be included in the dialogue context.
- Phenomenon: The dialogue interface cannot display image content such as test waveform diagrams and winding wiring diagrams in documents. Cause: Knowledge base image retrieval configuration is not enabled, or image files are not correctly embedded in document chunks, resulting in images failing to be retrieved and displayed properly.
- Phenomenon: The LLM response includes professional content outside the knowledge base, or custom models perform inconsistently between the dialogue application and workspace. Cause: Prompt engineering does not clearly limit usage to only technical documentation for this category, or model configurations for the workspace and dialogue application are not synchronized, resulting in differences in invoked model parameters.
How to Confirm Configuration is Complete
- Initiate a query targeting the winding process card for a single batch. Check that returned document content only includes information for the corresponding batch, with no redundant data from other batches. This verification process is compatible with the interface logic of FastGPT v4.17.0.
- Upload a test report document that includes images, initiate a query containing image descriptions, and check that the dialogue interface can normally display the relevant images.
- Check whether
variablesconfiguration has passed dedicated identifiers such as batch numbers and equipment numbers, confirming that the LLM can accurately associate corresponding documents. - Compare model configurations between the dialogue application and workspace, confirming that their retrieval rules and context parameters remain consistent.
The values provided are common starting points and should be measured against the reader's own samples.