What this category’s data looks like
Typical entities in this category are slaughterhouses, cooked meat and frozen prepared food enterprises. Daily documents include animal quarantine certificates, slaughter and cutting operation records, ingredient marination parameter records, thermal processing temperature logs, real-time cold chain temperature logs, and microbial laboratory test reports. Data sources include animal quarantine information systems, supplier batch ledgers, production traceability platforms, temperature collection equipment, metal detection data, and laboratory test results. Update cadence follows production batch cycles: quarantine and batch-related data are synced per production batch, production operation records are updated per shift, and cold chain temperature logs are continuously collected and updated. Most documents combine structured fields and numerical values. For example, cold chain temperature logs include collection time, station number, and real-time temperature value. Quarantine certificates include fixed fields such as batch number, quarantine identification number, and quarantine date.
What constraints do these characteristics impose on knowledge base retrieval and recall?
Data for this category is strongly tied to production batches. Retrieval must accurately match batch fields to avoid mistakenly recalling cross-batch documents. Otherwise, spoiled raw materials or non-compliant batches may enter production. High-volume, frequent cold chain temperature data creates many structurally duplicate documents. Retrieval must filter redundant results to prevent context overload. Compliance-related documents such as raw/cooked product zoning and microbial test reports require strict field matching. This prevents incorrect association of cross-contamination or non-compliant raw material documents. Errors directly impact product safety. Retrieved results must have high credibility to avoid irrelevant or incorrect documents entering the review process.
How to set configurations
| Configuration Item | Suggested Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches key compliance fields for meat processing, avoids recalling irrelevant batch documents, and ensures review accuracy |
limit | Top 8 entries | Covers core documents including quarantine, temperature, and test records, while controlling context length to avoid exceeding model processing limits |
usingReRank | Enabled | Improves relevance ranking accuracy for multi-source heterogeneous documents such as scanned quarantine certificates and CSV temperature logs |
searchMode | mixedRecall | Structured fields such as batch numbers and temperature values are suitable for keyword matching, while quarantine descriptions and test conclusions are suitable for vector retrieval |
collectionIds | Bind the knowledge base collection ID of the current production batch | Only recalls documents from the corresponding production batch, preventing irrelevant content from the full knowledge base from entering compliance reviews |
datasetSearchUsingExtensionQuery | Disabled | Production documents have fixed structures, so query expansion is unnecessary. This avoids introducing content unrelated to the current review |
Three common configuration mistakes
- Phenomenon: Retrieval results include quarantine certificate documents from non-current production batches. Cause: The
collectionIdsparameter is not configured to bind the knowledge base collection of the corresponding batch, resulting in recall of non-current batch documents from the full knowledge base. - Phenomenon: Multi-column Excel cold chain temperature logs are split across rows and columns during automatic chunking, making individual collection records unretrievable. Cause: The chunking unit is not specified as rows, and the default automatic chunking logic splits complete collection entries.
- Phenomenon: Knowledge base reference links are forcibly displayed in conversations, and empty references are returned even when no valid documents are matched. Cause: The reference display switch in the FastGPT v4.17.0 conversation configuration is not disabled, or the processing logic for empty retrieval results is not properly set.
How to confirm the configuration is correct
- Upload quarantine certificates and corresponding cold chain temperature logs for a single production batch, perform a retrieval operation, and check if returned results only include documents from that batch.
- Adjust the
similaritythreshold, and verify that the relevance of retrieval results meets the accuracy requirements for production reviews, with no irrelevant documents included. - Import multi-column Excel cold chain temperature logs, and confirm that chunking results show one independent collection record per row.
- In the FastGPT v4.17.0 backend configuration, check that the
usingReRankswitch andcollectionIdsbinding status match the preset configuration.
Note: The configuration values listed above are common starting points. Adjustments should be made based on the relevant samples and use cases of the deployment.