What Data for This Category Looks Like
Cosmetics production technical documents include formula sheets, weighing and feeding records, emulsification and filling records, cleaning and site clearance records, semi-finished and finished product inspection reports, label review forms, and more. Data sources include registered and filed formulas, raw material safety materials, production equipment, ERP/MES systems, laboratory data, and packaging material version files. Data updates are triggered with each production batch or line change. Each single document contains fields such as raw material names, restricted substance content, production parameters, batch numbers, packaging material claims, and more. Units cover common production units including weight, temperature, volume, and others.
Constraints on Vector Models and Indexes
This category has high data update frequency and clear trigger nodes. Indexes must support incremental synchronization to reduce synchronization time during production gaps. Documents have large structural differences across types, so different chunking strategies must be adapted to avoid splitting that damages the integrity of core fields such as formulas and inspection reports. Production data includes professional fields such as restricted substances, temperature, and batch numbers. Vector models must accurately encode text and numeric associated content with clear business meanings. It is also necessary to distinguish the index priorities of static filed data and dynamic production data to ensure recall efficiency for core review scenarios.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Ensures core field integrity for long documents such as formula sheets and inspection reports, avoids splitting that damages business logic |
| Overlap Length | 100–150 characters | Covers cross-chunk production parameter associations, such as consecutive weighing and feeding steps |
| maxToken | 8192 | Meets encoding requirements for complete raw material safety materials in filed formulas, FastGPT v4.17.0 supports this value range |
| defaultToken | 1024 | Meets conventional encoding length for single-batch production records, balances model computing power and recall accuracy |
| Similarity Threshold | 0.75–0.85 | Meets strict matching requirements for formula and filing consistency reviews in cosmetics production |
| Single Search Limit | Top 10 entries | Covers recall requirements for multiple document types, avoids missing key packaging material or raw material data |
Three Common Configuration Mistakes
- A 404 error occurs after configuring baidu-embedding-v1 through the UI. The cause is failure to correctly match the third-party embedding service's interface path and request headers, and failure to pass the correct API key parameters as required by the service.
- The m3e vector model cannot be added to the index after deployment in a non-GPU environment. The cause is failure to enable CPU mode for model operation, or insufficient memory allocated to the Docker container for loading the lightweight embedding model.
- The number of index results does not match expectations after incremental synchronization of production data. The cause is failure to correctly configure the incremental update trigger conditions, leading to repeated synchronization or missing cleaning and site clearance record data after line changes.
How to Verify Correct Configuration
- Upload a single typical formula sheet, check if the chunked text retains core raw material and production parameter fields, with no obvious splitting errors.
- Submit a formula and filing consistency review request, verify that the returned recall results include corresponding filed data and production records.
- Trigger an incremental synchronization after a line change, check that the index update log only synchronizes newly added cleaning and site clearance, feeding records, with no duplicate entries.
- View the vector model's resource usage, confirm that it can stably complete encoding tasks even in a non-GPU environment, with no obvious lag or errors.
The values provided are common starting points, and should be evaluated against the reader's own samples.