This configuration is formulated based on the standard logic of FastGPT v4.17.0.
What the data for this category looks like
This category mainly covers enterprises processing liquid milk, yogurt, milk powder and dairy ingredients. Daily core documents include raw milk acceptance records, standardized ingredient lists, pasteurization or UHT heat treatment records, CIP cleaning records, filling records, sample retention and microbial test reports. Data sources cover milk source detection systems, MES production systems, homogenization and heat treatment equipment, online temperature and flow collection modules, CIP systems, LIMS laboratory management systems. Data is updated in real time per single milk tank or single production batch. Fields include raw milk antibiotic and physicochemical indicators, heat treatment temperature and duration, CIP cleaning parameters, filling environment parameters, and more. Data formats vary across multiple systems.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source heterogeneous data sources require retrieval to adapt to both structured equipment-collected data and unstructured paper or system text records, and filter irrelevant documents by business domain. High-frequency real-time updated batch-specific data requires retrieval to support quick location of associated records by batch number and production time, to avoid interference from irrelevant cross-batch content during audits. The continuous production nature of perishable raw materials requires retrieval delay to match production rhythm, and cannot tolerate excessive waiting time. The accuracy of core audit indicators such as raw milk antibiotics and heat treatment parameters is directly linked to production safety, requiring retrieval recall precision to be higher than general scenarios, to avoid missing key information.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the precision requirements for core audit indicators such as raw milk acceptance and heat treatment, and avoids filtering key documents that meet business standards with low similarity |
limit | Top 6 entries | Covers multiple types of documents involved in a single production batch (raw milk, ingredients, heat treatment, CIP, etc.), while controlling the total context length to fit model input limits |
searchMode | mixedRecall | Balances structured temperature and flow data collected by equipment and text keywords from paper records, adapting to multi-source heterogeneous business data |
usingReRank | Enabled | Re-ranks retrieved multi-source documents by business priority, prioritizing high-risk audit items such as raw milk antibiotics and heat treatment parameters |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands the original query into search terms including batch number and production time, to match exclusive documents updated per batch |
collectionIds | Dedicated dairy processing collection ID | Isolates knowledge base documents from other categories, to avoid retrieving irrelevant cross-domain content |
Three common mistakes
- Empty retrieval results, with no matching documents shown in the interface or logs. The cause is failure to configure batch-based retrieval expansion, or setting the
similaritythreshold too high, which filters key indicator documents that meet business requirements. - In private deployment scenarios, passing a
limitparameter value greater than 6 during API calls results in no response. The cause is that the default retrieval upper limit configuration has not been adjusted, which restricts the number of documents retrieved per call. - An error is prompted when uploading production record documents, and files larger than 1 MB cannot be imported. The cause is that the platform's single-file upload threshold configuration has not been modified, and the default 1 MB limit is still in use.
How to confirm the configuration is correct
- Enter an audit question that includes the batch number of a single raw milk acceptance, and check whether the retrieved results include associated documents such as raw milk test reports and ingredient lists for that batch.
- Adjust the
similaritythreshold to different ranges, observe changes in the number of retrieval results, and confirm that the threshold configuration takes effect. - Call the API with a
limitparameter of 8, and check that the number of returned documents meets expectations. - Upload a heat treatment record document larger than 1 MB, and confirm that the import process has no errors and the document can be retrieved normally.
The values provided here are common starting points, and should be evaluated against the reader's own samples.