What Data Looks Like for This Category
Typical entities are slaughterhouses, cooked meat and frozen prepared food manufacturers. Daily documents include animal quarantine certificates, slaughter and cutting records, ingredient marinating records, thermal processing records, cold chain temperature records, and microbial test reports. Data sources cover animal quarantine information, supplier batch archives, production traceability ledgers, real-time temperature collection data, metal detection results, and laboratory test reports. Update frequency follows production batches and shifts. Cold chain temperature records are updated continuously in real time. Document fields include batch numbers, test times, temperature values, test results, and other items. There is no unified standardized format. Some handwritten records need to be converted to electronic documents.
Constraints Imposed on Source Citation and Traceability by These Characteristics
Traceability requirements for this category must accurately match production batches and quarantine information. Continuously updated cold chain temperature records require retrieval results to be locatable to specific time nodes. Mixed retrieval of multiple document types must avoid information confusion across batches and processes. Biological risk control requires traceability results to cover all key documents, and no quarantine certificates or temperature anomaly records may be omitted. Document fields vary significantly across different production links. Retrieval must accurately identify core identifiers such as batch numbers and times to avoid mismatching content from different workshops or raw/cooked batches.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 4000–6000 characters | Adapts to the content length of individual meat processing documents, avoids traceability confusion caused by cross-batch chunking |
quoteMaxToken | 1200–1500 tokens | Matches the information density of documents in this category, prevents key content from being truncated due to exceeding context limits |
responseData | Return only core business fields | Reduces redundant information, focuses on traceability-critical content such as batch numbers and test times |
detail | Enable full field return as needed | Retains all fields of all original documents when full traceability information is required |
| Retrieval Similarity Threshold | 0.75–0.85 | Accurately distinguishes records from different batches, avoids mismatching documents from raw/cooked batches or different workshops |
| Single Retrieval Recall Count | Top 8 results | Covers the retrieval needs of multiple associated documents per batch, ensures key information such as quarantine certificates and production records is not omitted |
This configuration is based on FastGPT v4.17.0.
Three Common Misconfigurations
- Phenomenon: Chunked content that exceeds the
quoteMaxTokenlimit is still included in the final citation. Cause: The chunk size is set larger than thequoteMaxTokenlimit, and no pre-truncation is applied to retrieval results. - Phenomenon: The number of results returned by the knowledge base citation merging module does not match the set single retrieval recall count. Cause: Irrelevant documents are not filtered by batch number, resulting in records from non-target batches being mixed into the results.
- Phenomenon: Retrieval results from different branches in a workflow cannot be aggregated uniformly. Cause: No unified batch identification field is added to the retrieval results of each branch, making it impossible for subsequent logic to identify grouping criteria.
How to Verify Correct Configuration
- Upload a single batch slaughter record document, enter a query containing the corresponding batch number, and check if the token count of the retrieved citation content matches the
quoteMaxTokensetting after testing. - Review the fields of the retrieval results to confirm that core business fields such as batch numbers and test times are included, matching the requirements of the
responseDataordetailconfiguration. - Trigger a workflow branch that retrieves from multiple knowledge bases, and confirm that results returned by each branch can be aggregated by batch number for unified logical processing.
- Adjust the retrieval similarity threshold, test with a query containing a similar batch number, and confirm that the returned results only include documents from the target batch.
The values listed above are common starting points, and should be evaluated against the reader's own samples.