What the data for this category looks like
Core data for this category comes from formula systems, raw material batch management modules, internal mixers, Mooney viscometers, rheometers, vulcanizing presses, and laboratory testing equipment. Daily generated documents include formula sheets, weighing records, mixing curves, rubber compound quick test reports, vulcanization process cards, mold usage records, and finished product performance reports. Data update rhythm changes with production nodes. New records are generated for each raw material batch, each batch of compound mixing, and each vulcanization cycle. Fields include process parameters, batch numbers, test values, and operation records. Some documents include visualized curve data.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
This category has high data update frequency and is tightly bound to production nodes. Retrieval and recall must accurately match parameters of the current production batch to avoid mixing cross-batch data. Multiple document types coexist, including structured process parameter tables and unstructured curve data. This requires adapting retrieval and recall logic for different content formats. The business impact of individual data is direct and clear. Incorrect recall will directly cause production deviations, so higher parameter matching accuracy is required for retrieval to avoid including irrelevant data in the context. Additionally, data volume grows with production scale. The number of recalled results must be controlled within a reasonable range to avoid context overload that affects model processing efficiency.
How to set the configurations
This configuration is based on FastGPT v4.17.0, and adapts to the knowledge base retrieval and recall needs of this scenario:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for process parameters in this category. Prevents introducing irrelevant data if set too low, and misses eligible matches if set too high |
limit | Top 8 entries | Controls context length, and adapts to the multi-parameter associated retrieval needs of mixing and vulcanization processes |
searchMode | embedding | Prioritizes matching semantic and numerical features of process parameters, and adapts to mixed retrieval of structured and unstructured documents |
usingReRank | Enabled | Performs secondary sorting on initial recall results to optimize matching accuracy for professional documents such as mixing curves and process cards |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on user questions, and covers retrieval scenarios for associated documents such as formula sheets and weighing records |
collectionIds | Exclusive collection IDs divided by production modules | Isolates knowledge base data from different workshops and batches, and avoids invalid recall across production units |
Three Common Misconfigurations
- Phenomenon: Retrieved process parameters do not match the current production batch, and the response content deviates significantly from the user’s professional question. Cause:
collectionIdsare not divided by production module or workshop, and knowledge base data from cross-production units is recalled, leading to parameter confusion. - Phenomenon: The number of retrieved results exceeds expectations, and excessive context length causes model processing delays. Cause: The
limitparameter is not set correctly, and too many irrelevant historical batch documents are recalled. - Phenomenon: English formula sheets and vulcanization process cards are not correctly recalled, and the model output language does not match the preset setting. Cause:
datasetSearchUsingExtensionQueryis not enabled, and no adapted extended queries are generated for professional documents, leading to failure of semantic matching.
How to Verify Proper Configuration
- Upload a single mixing curve document from a specified batch, initiate a query that includes parameters of that batch, and check whether the recalled results include the document.
- View retrieval logs to confirm that the matching threshold of the
similarityparameter falls within the preset range, with no matching results that are too low or too high. - Switch queries from different production workshops, and check whether the recalled results only include knowledge base documents under the corresponding
collectionIds. - After enabling
usingReRank, compare the initial recall results with the reranked results to confirm that professional process documents are sorted more closely to the user’s question.
The values provided are common starting points and should be measured against the user's own samples.