What this category's data looks like
Data for this category originates from registered or filed formulas, raw material safety documents, weighing and emulsification equipment, ERP/MES systems, laboratory records, and packaging material version files. Daily documents include formula work orders, weighing and feeding records, emulsification or filling records, cleaning and sanitization records, semi-finished and finished product inspection reports, and label review documents. Data updates trigger with each production batch and each line change. Document fields cover raw material component proportions, equipment operating parameter thresholds, batch numbers, filing numbers, and more. Units align with corresponding measurement standards based on document type.
What constraints do these characteristics impose on the "knowledge base retrieval and recall" link
Data updates for this category trigger with each production batch and each line change, requiring the knowledge base retrieval link to support high-frequency incremental updates and prioritize recalling the latest batch equipment parameters and process documents. Document types cover multiple categories including formulas, equipment records, inspection reports, and others. Document scenario weights must be distinguished during retrieval to avoid formula data interfering with accurate recall for equipment inspection. Fields include compliance-related content such as filing numbers. Recall results must match enterprise registered and filed information to prevent operational guidance that does not meet compliance requirements. Units for equipment parameters and inspection data vary by document type. The retrieval link must support contextually associated unit matching to avoid parameter confusion. Equipment inspection in production scenarios requires rapid response, so retrieval latency must be controlled within a reasonable range aligned with production rhythm.
How to set the configurations
These configurations are based on FastGPT v4.17.0, as detailed in the table below:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Cosmetics manufacturing equipment inspection requires precise matching of equipment parameters and compliance documents. This threshold balances recall rate and accuracy, avoiding interference from irrelevant formula data |
limit | Top 6–8 results | The context length for equipment inspection Q&A is limited. This value range avoids exceeding the large model's processing limit while covering sufficient equipment operation reference information |
searchMode | mixedRecall | Balances semantic matching and precise keyword matching for filing numbers, equipment numbers, and other identifiers, meeting the dual needs of compliance verification and parameter queries |
usingReRank | Enabled | Filters low-relevance recall results, improves the accuracy of equipment inspection Q&A, and adapts to retrieval scenarios for multiple document types |
datasetSearchUsingExtensionQuery | Disabled | The question scenarios for equipment inspection are clear, so no additional query expansion is needed, avoiding introducing content unrelated to production equipment |
collectionIds | Specify dedicated collection IDs for equipment inspection | Isolates documents from other categories or non-equipment-related content, ensuring retrieval results only cover knowledge base content for cosmetics manufacturing equipment |
Three common configuration mistakes
- Issue: Matching equipment documents can be found in the knowledge base backend, but no valid output is returned after passing to the large model. Cause: The
limitvalue is set too low, resulting in insufficient valid data recalled, or the context length exceeds the large model's processing limit, making the model unable to read the recalled content. - Issue: Manually created knowledge base documents are always attached to the root directory, and cannot generate a hierarchical structure using the preset parent ID. Cause: The parent collection ID corresponding to the
collectionIdsparameter is not correctly specified, or the parameter format does not meet FastGPT's verification rules. - Issue: Equipment inspection retrieval latency exceeds 10 seconds, with no obvious optimization even for single equipment document scenarios. Cause:
usingReRankis not enabled to filter redundant results, orsearchModeis configured for vector-only retrieval, adding unnecessary computation chains.
How to confirm the configuration is correctly applied
A test question for cosmetics manufacturing equipment inspection can be submitted, and recall results can be verified to include category-specific documents such as corresponding batch equipment parameters and compliance filing information. Retrieval chain logs can be checked to confirm that the actual values of configuration items such as similarity and limit match the preset configurations. An incremental update can be triggered to verify that newly added line change or batch documents can be normally retrieved and recalled. The large model's output content can be verified to confirm that it is generated solely based on the recalled knowledge base documents, with no external unrelated information.
The values provided are common starting points and should be measured against relevant operational samples.