What the data for this category looks like
Typical operators include bakeries, pastry producers, instant noodle factories and frozen pasta factories. Daily documents include formulas and feeding lists, production records, baking or steaming curves, metal detection records, cleaning and disinfection records, and factory inspection reports. Data sources cover raw and auxiliary material qualification certificates, ERP/MES, weighing systems, ovens or frying lines, laboratories, and food traceability ledgers. Update frequency is per batch, per shift, or daily. Document structures include structured tables, time-series curves, and text records, with fields covering raw material names, feeding quantities, allergen labels, baking parameters, shelf life, and other relevant items.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Baking and convenience food production lines undergo frequent changeovers. Formula and allergen information requires fast and accurate matching, so knowledge base retrieval must support multi-dimensional field matching. Multi-source heterogeneous data sources require the retrieval system to integrate documents of different formats, avoiding information silos. High update frequency means retrieval and recall must adapt to real-time or near-real-time data synchronization, ensuring the latest production records and inspection reports are used during audits. Allergen control is a core risk point. Retrieval must prioritize recalling feeding and cleaning documents related to allergens, to avoid mislabeling or cross-contamination. This category relies more on verification of formula and cleaning switching, so retrieval focus should be tilted toward cleaning and disinfection records and formula adjustment documents.
How to Configure the Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches key information such as allergen labels and formula parameters, balancing accuracy and recall coverage |
limit | Top 8 entries | Covers multiple types of documents for a single production batch, avoiding excessive results that interfere with auditor review |
usingReRank | Enabled | Re-ranks recalled results by relevance, prioritizing display of baking curves and allergen-related documents |
datasetSearchUsingExtensionQuery | Enabled | Optimizes retrieval queries for frequent line changeover scenarios, matching feeding lists and production records across different batches |
searchMode | mixedRecall | Balances precise field matching for formulas and text descriptions for production records, adapting to multiple document structures |
Three Common Mistakes
- Failing to select
text-embedding-ada-002when configuring the embedding model, or specifying another unconfigured model, triggers theundefined model must match "^(texterror. The cause is that the corresponding model has not been configured in the knowledge base settings, or the selected model falls outside the supported range of FastGPT v4.17.0. - Multi-column structured Excel files (such as feeding lists) imported after upload have chaotic automatic segmentation, and cannot split each row as an independent segment. The cause is that the segmentation rules have not been adjusted for multi-column data, and the default automatic segmentation fails to recognize the corresponding association between columns and rows.
- Local knowledge base documents are not uploaded, but knowledge base citation identifiers still appear in audit results. The cause is that the global knowledge base citation display switch has not been turned off, or irrelevant datasets have not been excluded from the current audit's retrieval scope.
How to Confirm the Configuration Is Correct
- Upload a test multi-column Excel feeding list, trigger automatic segmentation, and confirm that each row is split as an independent segment.
- Initiate a retrieval containing allergen keywords, and check whether the reference sources of the returned results only include the configured target datasets.
- Adjust the value of
similarity, and verify that the number of recalled results changes as expected based on the threshold. - Turn off the knowledge base citation display switch, initiate a retrieval, and confirm that no additional citation links or identifiers appear in the results.
The values provided here are common starting points, and should be evaluated against the reader's own samples.