What data for this category looks like
Data for this category comes from formula systems, raw material batch ledgers, internal mixer operation logs, Mooney viscosity and rheometer test data, vulcanizer process records, and laboratory finished product performance reports. Data updates follow the cadence of each raw material batch, each batch of compound rubber mixing, and each vulcanization cycle. Individual documents include fields such as formula component ratios, weighing and feeding details, mixing temperature curves, quick test indicators for compound rubber, vulcanization time and temperature parameters, mold number, and finished product tensile strength. Units include degrees Celsius, seconds, kilograms, megapascals, and others.
What constraints these characteristics impose on knowledge base retrieval and recall
First, high-frequency multi-dimensional data updates require the retrieval and recall process to support real-time matching of the latest batch parameters, preventing use of expired formula or process data. Second, multi-source heterogeneous document formats (curve, table, text) require the retrieval model to adapt to semantic matching of mixed fields, prioritizing recall of mixing or vulcanization parameters strongly relevant to the current production scenario. Third, individual documents have dense and highly correlated fields, so the number of recalled documents per single request must be limited to prevent context overload that interferes with the large model’s ability to generate accurate ledger reports. Fourth, the association logic between data from different stages (such as raw material batches corresponding to mixing processes) requires retrieval to support cross-document associated recall, covering parameter associations across the full production cycle.
How to set the configuration
This page’s configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Semantic similarity threshold for matching professional fields such as mixing processes and vulcanization parameters, to avoid recalling irrelevant general rubber documents |
limit | Top 6–8 entries | Individual mixing or vulcanization documents have many fields, limiting the number of recalled entries prevents context overflow |
usingReRank | Enabled | For multi-source heterogeneous documents, the re-ranking model can filter out parameter entries most relevant to the current production task (such as ledger summary), improving retrieval accuracy |
searchMode | mixedRecall | Balances keyword matching for formula components and semantic matching for process parameters, adapting to the retrieval needs of multiple fields in this category |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms related to production scenarios, such as expanding "mixing records" to "internal mixer temperature curve, rubber weighing record", covering associated data across multiple stages of this category |
collectionIds | Bind corresponding knowledge base collections by production stage | Bind the formula library, mixing log library, and vulcanization process library to their respective collections, narrowing the retrieval scope and improving efficiency |
Three common misconfigurations
- Symptom: After calling knowledge base retrieval,
usingReRankreturns false. Cause: The re-ranking model binding was not enabled in the FastGPT backend, or the re-ranking model configuration was not associated with the currently used knowledge base collection. - Symptom: When referencing variables in the knowledge base node of a workflow, there are no selectable values in the reference variable dropdown menu. Cause: Variables related to knowledge base IDs were not defined in global variables or workflow variables, or the variable type was not set to text.
- Symptom: When dynamically specifying a knowledge base, the reference variable dropdown menu is blank after selecting the variable to apply. Cause: The knowledge base collection ID configuration was not set as a referenceable global variable, or the variable scope did not cover the current workflow.
How to confirm the configuration is correct
- Manually enter a query term related to mixing processes, check whether the retrieval results include fields exclusive to this category such as internal mixer temperature and rubber compound ratio.
- After enabling
usingReRank, compare the relevance and number of retrieval results before and after enabling to confirm the re-ranking logic is working. - After configuring dynamic variables, enter the preset knowledge base ID variable in the workflow test panel, verify that the retrieval node can normally call the corresponding collection.
- Check the retrieval log to confirm that documents meeting the
similaritythreshold are recalled first, with no irrelevant data included.
The values given are common starting points and should be measured against the reader's own samples.