What the data for this category looks like
Core data for baked and convenience food production scenarios comes from daily business documents of bread and pastry, biscuit, instant noodle, and frozen pastry factories. These 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 systems, weighing equipment, ovens/frying lines, laboratory testing systems, and food trace ledgers. Data updates occur per batch, per shift, or daily. Document fields include batch numbers, feeding quantities, baking temperatures and durations, allergen labels, and more. Some documents include numerical curves and image assets.
What constraints these characteristics impose on knowledge base retrieval and recall
The high-frequency update rhythm of this category requires the retrieval pipeline to support incremental synchronization and real-time recall. This prevents ledger summary errors caused by delayed data.
Mixed document types (text, numerical curves, images) require retrieval to cover both semantic matching and structured field matching. Pure text semantic retrieval alone cannot meet this need. Frequent product line changes and allergen control requirements demand precise retrieval scope limiting. This avoids interference from irrelevant documents across batches or workshops.
Additionally, products are sold directly to consumers. Retrieval result accuracy is directly linked to food safety risks. Low-match results must be strictly controlled to avoid unnecessary recall.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Balances retrieval recall rate and precision. Prevents non-batch production records or cleaning documents from being included in results |
limit | Top 8 entries | Covers core fields of a single-batch production ledger. Controls context length to fit the context window for ledger summary and report generation |
searchMode | mixedRecall | Matches both text semantics and numerical features. Fits business documents with structured numerical data such as baking curves and feeding quantities |
usingReRank | Enabled | Filters low-relevance recall results. Prioritizes high-value content related to allergen control and batch verification |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval term scope. Adds matching for fields such as batch numbers, workshop numbers, and product types. Improves retrieval precision |
collectionIds | Grouped by workshop/product line | Limits the knowledge base collections targeted by retrieval. Avoids interference from documents across batches or product lines |
Three common mistakes
- After uploading formula or production record documents that include images, retrieval results do not return image content. Cause: Knowledge base image parsing configuration is not enabled. Only text content within the documents is extracted.
- After upgrading to FastGPT v4.17.0, creating a knowledge base with a URL fails. Background logs display
cannot fetch internal url. Cause: Internal business system URLs are not added to the allowed access whitelist, or network proxy configurations do not cover internal interface requests. - After deploying the open-source version of FastGPT locally, the number of createable knowledge bases cannot be adjusted above the preset upper limit. Cause: The default knowledge base quantity limit parameter in the configuration file has not been modified.
How to confirm configurations are set correctly
- Upload single-batch formula and production record documents. Submit a retrieval request that includes the batch number keyword. Verify that the number of returned results matches the
limitconfiguration. - Enter a retrieval term that includes an allergen keyword. Confirm that returned results include cleaning and disinfection records and allergen control documents for the corresponding batch.
- Check background retrieval logs. Confirm that
usingReRankanddatasetSearchUsingExtensionQueryare executing normally. - Upload a document that includes images. Submit a retrieval request. Confirm that results include image links or parsed image description text.
The values provided here are common starting points. They should be evaluated against the reader's own samples.