What the data for this category looks like
Data for this category comes from animal quarantine information, supplier batches, production traceability, temperature collection, metal detection, and laboratory records. Daily documents include quarantine certificates, slaughter and cutting records, ingredient marinating records, thermal processing records, cold chain temperature records, and microbial test reports. Data updates run per batch and per shift, with continuous temperature control data synced in real time. Document fields include batch number, quarantine date, temperature value, test results, and more. Cold chain records use degrees Celsius as the unit. Slaughter records mark station numbers and carcass weight.
What constraints these characteristics impose on knowledge base retrieval and recall
This category has strong batch association attributes. Retrieval must accurately match the batch number field to avoid interference from cross-batch documents. Continuous temperature control data is generated in real time and has a large volume, so retrieval must support fast recall of high-frequency data. Separate raw and cooked product production requires retrieval results to be filtered by production area, preventing confusion of documents related to cross-contamination. Quarantine certificates and production records are tightly bound, so retrieval must associate compliant documents for the corresponding batch to ensure accurate compliance basis for bid quotations. Data update frequency is high, so the retrieval link must synchronize the latest production and test data to avoid outdated bid reference materials.
How to configure the system
This page’s configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches strongly associated fields such as quarantine certificates and batch records, avoiding recall of irrelevant low-matching documents |
limit | Top 12 entries | Covers retrieval needs for multiple document types per batch, while controlling context length to fit large model input limits |
searchMode | mixedRecall | Combines semantic and keyword retrieval, adapting to structured fields like batch numbers and temperature values, as well as compliance description text retrieval |
usingReRank | Enabled | Filters redundant results from semantic retrieval, accurately matching compliance basis and production data required for bids |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval scope to cover associated documents for supplier batches and production traceability, improving completeness of compliance and production basis for bid quotations |
Three common configuration mistakes
- Online retrieval takes more than 10 seconds. This happens when the
limitparameter value is not restricted, recalling too many unnecessary documents and causing excessive embedding and reranking calculation load. - Batch number search results are irrelevant. This occurs when only pure semantic retrieval mode is used, without enabling
searchModeas hybrid. Pure semantic retrieval cannot accurately match structured batch number fields, leading to recall of irrelevant documents. - After setting
limitto 3000, the large model does not receive context content. This is caused by failing to adjust the total context length threshold, exceeding the large model’s input limit and truncating the context, making normal transmission impossible.
How to confirm the configuration is correctly set
- View retrieval logs to confirm that
similarityscores of returned results fall within the set range, with no documents below the threshold recalled. - Enter batch number keywords to verify that when
searchModeis set to hybrid, quarantine and production records for the corresponding batch are accurately recalled. - After enabling
usingReRank, compare the order of recall results before and after enabling to confirm that high-matching documents are returned first. - Adjust the
limitparameter to a reasonable range, verifying that the large model can normally receive the returned context content with no truncation errors.
The values provided here are common starting points and should be measured against the reader's own samples.