What the data for this category looks like
This category covers the full production process of factories producing carbonated drinks, juice, tea, functional beverages, and packaged drinking water. Daily documents include formula sheets, water treatment records, blending records, sterilization or aseptic filling records, packaging material inspection forms, and finished product inspection forms. Data sources include water source and raw material testing, formula systems, online sugar and acidity monitoring equipment, filling equipment, visual inspection systems, and LIMS systems. Updates are performed by production batch, per shift, or through continuous collection. Each record is linked to a unique production batch number, and includes fields such as batch number, timestamp, test values, and equipment number.
Constraints on Knowledge Base Retrieval and Recall
Structured data archived by batch requires retrieval to prioritize matching the batch number field, to avoid interference from irrelevant cross-batch results. Frequently updated continuous collection data requires the knowledge base to support incremental synchronization, to prevent data lag. Diverse documents from multi-category switching require retrieval to cover associated records for different formulas and equipment parameters. A single batch is typically linked to 5 to 15 test records. This requires recall results to be aggregated by batch, to avoid scattered individual entries. Additionally, precise comparison of test values requires retrieval to balance semantic matching and precise matching logic for numerical fields.
Configuration Settings
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match precise fields such as batch numbers, while covering semantic associations for formulas and test records |
limit | Top 10–15 entries | A single batch links to multiple records; balances recall coverage and retrieval response speed |
searchMode | mixedRecall | Balance semantic matching and precise matching for keyword fields such as batch numbers and equipment numbers |
usingReRank | Enabled | Filter redundant frequently updated test records, focus on documents directly related to quality issues |
datasetSearchUsingExtensionQuery | Enabled | Automatically expand keywords related to quality issues, covering formula terminology for different beverage categories |
collectionIds | Specify knowledge base collections bound to production batches | Limit retrieval scope, avoid interference from irrelevant cross-batch data on recall results |
Three Common Configuration Mistakes
- Retrieval results mix documents from multiple production batches, and some requests return permission denied prompts. Cause: A universal authentication key was used when calling the knowledge base API, and the knowledge base collection corresponding to the target production batch was not bound. This causes the retrieval scope to exceed preset permissions.
- Training errors occur when importing modified CSV knowledge base data, and the system returns a field format mismatch prompt. Cause: Required fields such as batch number and test values were not adjusted according to the CSV template for the target FastGPT version. For example, old format data exported from v4.9.2 was directly imported into v4.12.1.
- Only a small number of results are returned when retrieving high-batch data, and all test records for the current batch cannot be covered. Cause: The reference upper limit of the locally deployed knowledge base search module was not adjusted according to the data update frequency, resulting in result truncation.
How to Verify Successful Configuration
- Enter the target production batch number as the retrieval keyword, verify that returned results only include documents linked to this batch.
- Trigger an incremental synchronization, verify that the knowledge base has updated the latest water treatment and filling records by production batch.
- View the retrieval log, confirm that
usingReRankanddatasetSearchUsingExtensionQueryare active, with no redundant results. - Adjust the
similarityparameter, verify that recall results under different thresholds meet expected matching accuracy.
The values provided are common starting points and should be evaluated against the reader's own samples.