What the data for this category looks like
This category covers daily production documents for carbonated drink, juice, tea, functional beverage, and packaged drinking water factories. Documents include formula sheets, water treatment records, blending records, sterilization or aseptic filling records, packaging material inspection reports, and finished product inspection reports. Data comes from water source and raw material testing, formula systems, online sugar and acidity monitoring equipment, filling equipment, visual inspection systems, and LIMS systems. Data is updated per batch, per shift, or via continuous collection. Document structures include fields such as batch number, test values, and operating parameters. Some records include equipment operating parameters and test result units.
Constraints Imposed on Knowledge Base Retrieval and Recall
Production data for this category updates frequently and is tied to batch information. This requires knowledge base retrieval to accurately match batch dimensions, to avoid mixing cross-batch data. Multiple heterogeneous data sources create large differences in document formats. Support for unified recall of multiple text types is needed. Frequent product switching leads to large volumes of new documents. Adaptation to high-frequency updated dataset synchronization is required. Exception reports cover specific scenarios such as filling liquid level, sealing, and label batch number. Retrieval must cover associated documents for specific equipment and links. At the same time, control the relevance of returned results to avoid introducing irrelevant production data from other categories.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precise parameter requirements for beverage production, and avoids low-relevance water treatment records from being included in formula retrieval results |
limit | Top 5–7 entries | Covers associated documents for a single batch production, while controlling context length |
usingReRank | Enabled | Optimizes the sorting of retrieval results for filling exception reports, prioritizing matches for high-frequency exception scenarios such as sealing and liquid level |
datasetSearchUsingExtensionQuery | Enabled | Supplements exclusive fields such as batch number and equipment model to improve retrieval matching accuracy |
collectionIds | Bound to exclusive beverage production dataset ID | Excludes knowledge base data from other categories to avoid mixed retrieval results |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
Three Common Misconfigurations
- Phenomenon: Retrieval results split grouped chapter titles, links, and summaries. Cause: The
datasetSearchUsingExtensionQueryparameter was not configured correctly, leading to grouped documents being recalled individually. - Phenomenon: Returned results include production documents from non-beverage categories. Cause: Exclusive
collectionIdswere not bound, so the retrieval scope covered knowledge base datasets from other categories. - Phenomenon: A parameter error is returned when calling the knowledge base interface to create a training order. Cause: The optional value of
trainingTypewas not specified correctly, or thedata-rawfield content was not passed as required.
How to Verify Successful Configuration
- Upload a set of test documents that include chapter titles, links, and summaries. Run a retrieval, then confirm that all three elements are recalled simultaneously and not split.
- Search for production keywords from non-beverage categories. Confirm that no relevant content appears in the returned results.
- Call the knowledge base interface to create a training order. Pass a valid
trainingTypeparameter, then confirm that the returned status code is 200. - Adjust the
similarityparameter range. Search for precise parameter keywords, then confirm that the relevance of the returned results meets expectations.
The values provided are common starting points and should be measured against the reader's own samples.