What this category’s data includes
Data for this category comes from daily business documents of skincare, haircare, makeup, and special cosmetics manufacturers. These documents include formula work orders, weighing and feeding records, emulsification and filling records, cleaning and site clearance records, semi-finished and finished product inspection reports, and label review documents. Data sources include registered and filed formulas, raw material safety materials, production equipment, ERP/MES systems, laboratory data, and packaging material version files. Updates trigger per production batch or each line change. Documents contain fields such as formula components, operating parameters, test results, and packaging material information. Matching between filed information and actual production data is required.
Constraints on knowledge base retrieval and recall
The high-frequency updates, mixed document types, and strict compliance matching requirements of this category create multiple constraints for knowledge base retrieval and recall. The per-batch or line-change update frequency requires the retrieval system to support incremental synchronization and rapid refresh, to avoid using expired data. Multiple document types cover different fields including process parameters, test results, and filed information. This requires precise cross-type document matching retrieval. The need to verify consistency between formulas and filed information requires retrieval results to link to filed data sources, to ensure matching accuracy. Real-time operating verification during production requires balancing response efficiency for both long document retrieval and short parameter queries.
The following configurations are based on FastGPT v4.17.0.
Configuration settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | For strict compliance scenarios such as formula and filing matching, to ensure retrieval result relevance meets standards |
limit | Top 8 entries | Covers core audit documents for a single production batch, to avoid excessive redundant results that cause interference |
usingReRank | Enabled | Optimizes sorting of multiple document types, prioritizing results related to formula verification and raw material checks |
datasetSearchUsingExtensionQuery | Enabled | Optimizes semantic matching for colloquial queries during production operations |
searchMode | mixedRecall | Meets retrieval needs for multiple data sources including process documents, inspection records, and filing data |
collectionIds | Knowledge base collections linked to production batches | Separates documents from different production workshops or product lines, to avoid interference from cross-batch data |
Three common configuration errors
- Issue: Document chunks are lost after setting splitting parameters. Cause: The splitting logic for long cosmetics manufacturing documents (such as complete formula work orders) was not adapted, leading to failure to correctly split long text.
- Issue: Rate limit errors occur during vectorization. Cause: Concurrency processing configurations were not adjusted based on the total data volume of documents for a single production batch, leading to requests exceeding rate limits.
- Issue: Retrieval result latency exceeds expected levels. Cause: Reranking function was not enabled, or retrieval mode was not optimized, leading to a large number of irrelevant documents participating in retrieval calculations.
How to verify correct configuration
- Upload typical documents for a single batch (such as a complete formula work order), check that complete document chunks are generated with no missing content.
- Run a formula compliance retrieval query, confirm that the similarity range of returned results matches the preset configuration.
- Simulate high-frequency retrieval requests, confirm that response time meets the real-time requirements of business audits.
- Switch between knowledge base collections for different product lines, confirm that retrieval results only link to the currently configured dataset.
The values provided are common starting points and should be measured against the reader's own samples.