What the data for this category looks like
This category's data mainly comes from daily production and compliance documents of bakeries, cookie, instant noodle, and frozen pastry factories. Documents include formulas and ingredient lists, production records, baking/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 systems, production equipment, laboratories, and food traceability ledgers. Update frequency is per batch, per shift, or daily. Most documents contain precise formula parameters, production timing data, and compliance testing fields.
What Constraints These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Multi-batch, high-frequency updated data for this category requires the knowledge base to support real-time incremental recall. Delayed data can cause mismatches between bid quotations and actual production, especially in key links such as formula review, allergen materials, and production windows. Mixed multi-type document structures require the retrieval link to adapt to field extraction for different formats, including formula parameters, production curves, and metal detection records. The ready-to-eat product attribute means compliance content in bid documents must accurately match testing records. Retrieval results must strictly link to production data for corresponding batches. They must also quickly distinguish formula differences between products that have undergone line changes, to avoid cross-category confusion and ensure compliance and accuracy of bid content.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets retrieval requirements for precise parameters such as allergens and production windows |
limit | Top 8 entries | Covers multiple types of document data for a single production batch |
searchMode | embedding | Adapts to semantic matching of unstructured text such as formulas and production curves |
usingReRank | Enabled | Performs secondary sorting of initial recall results by relevance, improving recall accuracy for key parameters |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval scope to cover associated information of formula ingredient lists and production records |
collectionIds | Specify the knowledge base collection ID for the target category | Limits retrieval scope to exclude document data from unrelated categories |
Three Common Mistakes
- Issue: Retrieval results do not include image information in docx documents, making it impossible to extract formula parameters from images. Cause: Image recognition configuration for the knowledge base is not enabled, only text content is extracted.
- Issue: The fixed phrase "FastGPT is a knowledge base question answering system based on large language models (LLM)" always appears at the end of test responses. Cause: The default system appended text is not disabled in the system prompt, or the prompt does not include shielding rules for this content.
- Issue: The number of documents returned by retrieval does not match the configured
limitparameter, or documents from unrelated categories appear. Cause: ThecollectionIdsconfiguration for limiting retrieval collections is not set correctly, or thesimilaritythreshold is set too low, leading to recall of irrelevant results.
How to Confirm the Configuration is Correct
- Upload a docx formula document that includes images, perform retrieval in the FastGPT v4.17.0 environment, and verify that the results include the parameter information from the images.
- Initiate a test question and answer, check whether fixed system appended text appears at the end of the response, and confirm that irrelevant content has been shielded.
- Adjust the
similaritythreshold, observe changes in the relevance of retrieval results, and verify that the parameter takes effect. - View the knowledge base retrieval logs, confirm that only document data within the specified collection is returned, and verify that the
collectionIdsconfiguration is correct.
The values provided here are common starting points and should be measured against the reader's own samples.