What Data for This Category Looks Like
This category of data is generated from sources including PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates are triggered irregularly with each change. Daily documents include ECN/ECR, change assessments, drawings and BOM versions, process validation reports, transition plans, inventory disposal forms, and customer approval documents. Fields include change number, effective node, associated materials/processes, validation status, and others. There is no unified fixed cycle; each change triggers the addition or update of corresponding data.
Constraints Imposed by These Characteristics on Multi-turn Dialogue and Prompt Engineering
Scattered data sources require linking to multiple knowledge base fragments. Prompts must explicitly specify calling documents from the correct sources. Irregular updates require the dialogue system to support on-demand knowledge base refreshes, preventing document expiration caused by fixed-cycle synchronization. Documents include version numbers and associated materials or process fields. Multi-turn dialogue must track version links for changes to avoid mixing outdated files. Changes involve cross-department collaboration, so multi-turn dialogue must preserve the messages context to carry forward change requests from different roles. This ensures drafted documents cover all associated impacts.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Single Search Limit | Top 6 results | This category's document count increases with changes, so limit returned results to avoid context overload |
| Reference Content Template | Change Number: {Number}, Effective Time: {Time}, Associated Process: {Process} | Matches core fields of this category's documents, enabling quick extraction of key change information |
| Maximum Context Length | 8000–12000 tokens | Adapts to multi-turn dialogue needs for retaining cross-department change requests and associated document content |
| Variable Configuration | variables binds change number and effective node | Used to quickly locate specific change documents in the knowledge base |
| Conversation ID Binding | chatId binds the current change session | Used to distinguish dialogue contexts for different changes to avoid session confusion. FastGPT v4.17.0 and later support this configuration |
| Query Optimization | Enabled | Helps the model convert vague change requests into precise knowledge base retrieval instructions |
| Reranking | Enabled | Prioritizes returning document fragments most relevant to the current change topic |
Three Common Mistakes
- Retrieval results include irrelevant outdated change documents. This occurs because the prompt does not explicitly specify linking to the specified files for the current change, causing the model to retrieve mismatched historical files.
- The dialogue cannot call the specified newly created change knowledge base. This occurs because the corresponding change-specific variable is not bound in the configuration, so the system cannot locate the correct knowledge base link.
- Drafted documents mix content from the prompt and knowledge base reference templates. This occurs because the distinction between the prompt's instruction logic and the knowledge base reference's content format is not clearly defined, leading to chaotic model output.
How to Verify Proper Configuration
- Enter a simulated change request, check if the retrieved documents include the current change's number and associated process, and verify field matching.
- Initiate a multi-turn dialogue, submit change supplement requests from different departments consecutively, check if the context is preserved and correctly handled by the model.
- View the variable binding section in the configuration interface, confirm that core fields related to changes have been correctly linked to the dialogue workflow.
- Trigger a knowledge base refresh operation, check if newly uploaded change documents can be properly retrieved in subsequent searches.
The values listed are common starting points and should be evaluated against the user's own samples.