What the Data for This Category Looks Like
Core data for cosmetics production comes from registered or filed formulas, raw material safety documents, ERP/MES systems, laboratory equipment, and packaging material versions. Data is updated with every production batch and every line change. Daily audit documents include formula work orders, weighing and feeding records, emulsification or filling records, cleaning and site clearance records, semi-finished and finished product inspection reports, and label review forms. Data fields include restricted substance content, feeding weight, emulsification temperature, batch number, packaging material version, and other items, with corresponding units such as kilograms, degrees Celsius, and batch numbers.
Constraints for Multi-turn Dialogue and Prompt Engineering
Since data is updated with every production batch and line change, multi-turn dialogue must bind the exclusive context of the current batch to prevent cross-batch data mixing. Audits cover multiple areas including formula compliance, process control, and label claims. Prompts must be split by process to guide dialogue through the sequence of formula verification, feeding inspection, process confirmation, and inspection review. Documents include highly regulated fields such as restricted substances and filed formulas. Prompts must forcibly link uploaded filed data to ensure audit results align with registration requirements. Data sources are dispersed. Dialogue must support dynamic calls to different document types to match the needs of the current audit step.
Configuration Settings
This configuration applies to FastGPT v4.17.0
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
variables | Enabled, bind current batch dynamic fields | Quickly substitute data such as feeding quantity and batch number to reduce repeated input |
chatId | Bind current production batch ID | Associate exclusive conversation records to avoid mixing data across batches |
messages | Limit context to the first 20 conversation turns or 8000–12000 characters, calibrated based on actual testing | Match the multi-link audit process of each production batch to avoid context overflow interfering with current batch judgment |
| Single search limit | Top 6 matching results | Cover multiple types of documents including raw materials, processes, labels, etc., to avoid interference from redundant information |
| Question optimization | Enabled, split prompts by production link | Correspond to audit steps of formula, feeding, emulsification, and inspection to guide dialogue to proceed in order |
Three Common Misconfigurations
- Phenomenon: Conversation records are lost after refreshing the conversation page, and a new conversation is displayed. Cause: The
chatIdparameter is not correctly bound, and conversation context is not stored in association with the current production batch. - Phenomenon: The AI prompts that no files are uploaded, but the system prompt has been associated with formula documents. Cause: Knowledge base chunking does not match the field structure of cosmetics production documents, so uploaded files such as filed formulas and feeding records cannot be recognized.
- Phenomenon: Custom Word download links cannot be generated during the conversation. Cause: Dynamic data such as current batch number and feeding quantity are not passed through the
variablesparameter, so the workflow cannot obtain valid content to generate files.
How to Verify Proper Configuration
- Initiate an audit conversation for the corresponding batch, and check whether the context includes dynamic data such as current batch feeding and process information.
- Refresh the conversation page, confirm that conversation records are not lost, and the session identifier matches the current production batch.
- Upload a filed formula document, initiate a query about formula compliance, and confirm that the AI can correctly read the document content.
- Trigger the workflow export function, and confirm that the generated file includes core audit fields of the current batch.
The values listed here are common starting points and should be evaluated against the samples relevant to the specific deployment.