What this category’s data looks like
Cosmetics marketing data primarily comes from official filing systems of beauty brands partnered with financial institutions, public databases of ingredient suppliers, compliant product detail pages on e-commerce platforms, and exclusive marketing materials provided by partner brands. Update frequency changes dynamically alongside new product launches, regulatory policy adjustments, and partner marketing campaigns. Each updated document follows a uniform structure, including product identifier, filing number, ingredient list with concentration markings, skin type classification, efficacy description, compliance notice, packaging specifications, and pricing units. Most data fields include clear unit markers, such as percentage for ingredient concentration, milliliter or gram for packaging volume, and yuan for price.
Constraints imposed by these characteristics on model integration and configuration
Cosmetics category data includes ingredient fields with concentration markings, fixed enumerated skin type classifications, mandatory compliance notice text, and updates that adjust dynamically with new product launches and partner marketing campaigns. These characteristics require sufficient context length to be configured during model integration, to ensure that numerical values with units such as ingredient concentrations are fully parsed. Compliance notice text is mandatory for financial institution marketing, so recall priority must be configured to ensure it is prioritized in the context. Dynamically updated data sources require a scheduled synchronization mechanism to avoid using expired filing data. For fixed enumerated skin type classification fields, field weights for vector recall must be configured to prioritize matching core business-related marketing scenarios.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Accommodate complete cosmetics product documentation, ensure long text such as ingredients and compliance notices is fully included in the context, and comply with FastGPT v4.8 and later configuration specifications |
llmModels | Match the model identifiers bound to OneAPI exactly | Avoid loading failures caused by mismatched model identifiers, ensure correct routing to the target model during calls, and comply with FastGPT v4.8 and later configuration formats |
ONEAPI_API_BASE | Match the deployed OneAPI service address exactly | Ensure model requests are sent correctly to the specified transit service, avoid routing errors |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Cover the parsing time required for cosmetics filing documents, avoid parsing task timeouts caused by large document content |
recallTopK | Top 8–12 entries | Balance recall relevance and coverage, ensure marketing content covers multiple dimensions such as ingredients, efficacy, and compliance |
similarityThreshold | 0.75–0.85 | Filter low-relevance recall results while retaining necessary low-similarity text such as compliance notices |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Issue: After configuring multiple models, only the first model in the
llmModelsarray is used during calls, or a model identifier mismatch error is displayed. Cause: The model identifiers in thellmModelsarray are not fully aligned with the return identifiers of OneAPI or model service providers, or all available models are not correctly declared in the configuration file. - Issue: Knowledge base responses are truncated, and complete product ingredient or compliance notice content cannot be displayed. Cause: The
maxContextconfiguration value is too small to accommodate complete cosmetics product documentation, resulting in early context truncation. - Issue: OneAPI interface verification passes, but no model response is returned after initiating a query. Cause: The interface path suffix is omitted from the
ONEAPI_API_BASEconfiguration, or the model specified inllmModelshas not been mounted in the OneAPI service.
How to verify successful configuration
- Access the model management page in the FastGPT backend, and confirm that the configured model list exactly matches the identifiers in the
llmModelsarray. - Upload a single complete cosmetics product filing document, run a parsing test, and confirm that the parsing time does not exceed the configured
PARSE_FILE_TIMEOUT_SECONDSvalue. - Initiate a test query containing the keywords "ingredient concentration" and "compliance notice", and check whether the returned results include complete relevant field content.
- Review the OneAPI interface call logs, and confirm that requests initiated by FastGPT are correctly routed to the target model service.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.