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. Daily documents include formula sheets, weighing records, mixing curve charts, rubber compound quick test forms, vulcanization process cards, mold usage records, and finished product performance test reports. Data updates occur for every raw material batch delivery, every rubber compound production run, and every product forming cycle per mold. Document fields include raw rubber grade, proportion share, mixing temperature range, vulcanization time, mold number, and more. Corresponding units include degrees Celsius, minutes, MPa, and other professional industrial parameters.
What constraints these characteristics impose on knowledge base retrieval and recall
The high-frequency update nature of this category requires the knowledge base to support incremental synchronization, to avoid audit deviations caused by delayed data. Multiple document types (text process cards, tabular records, curve charts) require retrieval capabilities to cover content parsing and matching across different formats. The presence of professional fields and units requires retrieval to support precise field matching, to avoid irrelevant data being recalled. The independent data attribute of individual batches or mold cycles requires recall results to be strongly associated with the current audit batch, and data must not be mixed across batches. This page provides configuration guidance based on FastGPT v4.17.0.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.88 | Meets the precision requirement for matching professional fields of mixing processes and vulcanization parameters, avoids mistakenly recalling irrelevant rubber compound data |
limit | Top 8–12 entries | Covers all associated formula, process, and inspection data required for a single audit, avoids excessive results interfering with judgment |
searchMode | mixedRecall | Adapts to multi-type document retrieval including mixing curves (chart data), process cards (text), and weighing records (structured tables) |
usingReRank | Enabled | Reorders recall results by business relevance, prioritizes matching the raw material batch information of the current batch |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval keywords, supplements retrieval coverage for professional fields such as raw material batch numbers and mold numbers |
collectionIds | Bind the dedicated knowledge base collection for the current category | Isolates documents from other categories, avoids recalling irrelevant data outside tire production |
Three common errors
- When attempting to access a mixing curve XLSX file stored in the knowledge base during a conversation, the system returns that the file content cannot be read. This occurs when the parsing plugin for tabular documents is not enabled, or file content extraction was not correctly enabled during upload.
- Retrieval results include rubber compound formula data from non-current batches. This happens if the
similaritythreshold is set too low, or the dedicatedcollectionIdsis not bound, leading to cross-collection recall of irrelevant data. - The number of returned retrieval results is far lower than expected. This occurs if the
limitvalue is too small, orsearchModedoes not support structured table field retrieval logic, resulting in professional fields within tables not being effectively matched.
How to verify correct configuration
- Upload a standard mixing XLSX file to the corresponding knowledge base, initiate a retrieval containing the phrase "weighing record in row 3", and confirm that specific parameter content from the table can be extracted.
- Enter the raw material batch number of the current batch, confirm that retrieval results prioritize returning the corresponding batch’s formula sheet and weighing record, to verify the effectiveness of batch association matching.
- Adjust the
similaritythreshold, confirm that a lower threshold will recall more associated data, while a higher threshold will only return strongly matching process data. - Check the binding status of the knowledge base collection, confirm that only the document collection for the current category is associated, and no cross-category documents are included in recall results.
The values provided here are common starting points and should be evaluated against the reader's own samples.