What This Category of Data Looks Like
Data sources for this category include PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates are triggered irregularly with each process or engineering change. Daily documents include ECN/ECR forms, change assessment reports, process drawings and BOM version files, process validation reports, production switching plans, inventory disposal documents, and customer approval documents. Data fields match corresponding business dimensions based on change scenarios, including change number, version identifier, effective node, involved process scope, associated material code, and more. No unified fixed general statistical fields exist.
Constraints Imposed by These Characteristics on the Multi-turn Dialogue and Prompt Words Link
Data sources are scattered and updated dynamically with changes. Multi-turn dialogue must support cross-system associated data queries. Prompt words must clearly limit retrieval scope to the document set associated with the current change. Document types are diverse and fields have no unified standards. Prompt words must guide large models to prioritize matching core identifiers such as change numbers and version numbers, to avoid confusing different batches of change data. Change triggers have no fixed cycle. Dialogue context must support temporary loading of newly uploaded change files. The number of documents retrieved in a single dialogue must be limited to prevent redundant information from interfering. Changes involve cross-department collaboration. Multi-turn dialogue must retain historical interactive change context, to ensure subsequent questions can be linked to previous validation results or approval statuses.
Configuration Settings
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Maximum Context | Top 20000 tokens | Adapt to the length of single change documents such as ECN/ECR and process validation reports, to avoid exceeding the model's context limit |
| Overlap Length | 200-300 characters | Retain contextual association of change documents, avoid splitting key paragraphs such as version switching instructions and process associations |
| Single Search Limit | Top 8 entries | Match the number of multi-type documents associated with each change, prevent redundant information from interfering with core change logic |
| Question Optimization | Enabled | Help refine professional terms and core identifiers related to changes, avoid retrieval deviations caused by vague questions |
variables | Bind change number and version number variables | Accurately locate change documents associated with the current session, avoid confusing different batches of change data |
messages | Retain the last 10 rounds of interactions | Store change dialogue history for cross-department collaboration, ensure subsequent questions can be linked to previous approval statuses |
chatId | Bind session-specific ID | Retain historical interactive change context, ensure cross-round questions can be linked to previous validation or approval information |
Three Common Misconfigurations
- Symptom: A connection error prompt appears after initiating a dialogue related to process changes, and the issue cannot be resolved after enabling a network proxy. Cause: Network parameters of the model interface are not configured correctly, and proxy rules do not cover the full link of model requests, resulting in failure to connect to the target model service.
- Symptom: Process drawings and validation report images returned by the knowledge base cannot be displayed normally in the dialogue interface. Cause: Prompt words do not explicitly require the large model to adapt to image display formats, and the associated binding settings for knowledge base image retrieval are not configured, resulting in the large model only returning text descriptions.
- Symptom: Keyword variables such as change effective date and associated material code in the dialogue cannot be accurately extracted. Cause: The
variablesparameter is not used to bind core change identifier fields, and prompt words do not explicitly specify the variable types to be extracted, resulting in the large model being unable to accurately match target information.
How to Verify Proper Configuration
- Initiate a query containing a specified change number. Verify that returned documents are only associated with business files of the corresponding change.
- Initiate two or more consecutive queries. Verify that historical interactive change context is retained, and subsequent queries can be linked to previous validation or approval information.
- Upload the latest change-related documents. Initiate an associated query. Verify that the dialogue can retrieve the content of the newly uploaded files.
- Test a query involving image association. Verify that the large model can correctly process image-related content in the knowledge base.
The values provided are common starting points and should be evaluated against relevant internal samples.