What this category of data includes
Data for this category comes from enterprise MES systems, online inspection equipment, laboratory test records, process parameter documents, customer quality standards, and historical defect databases. Update cadence follows per production batch, per shift synchronization, or continuous real-time collection of test data. Daily documents include control plans, inspection records, SPC reports, finished product inspection reports, nonconforming product review forms, release forms, and monthly quality reports. Document fields include inspection items, tolerance ranges, inspection frequencies, batch identifiers, conformity judgment items, and more. Some documents are linked to specific production equipment numbers and process parameters.
Constraints for the Knowledge Base Retrieval and Recall Link
Data sources for this category are scattered, and update cadences vary significantly. Some data is real-time collected test data, while some is batch-synced production documents. This places requirements on retrieval real-time performance and incremental update capabilities. Diverse document formats include structured reports and long-text specifications, so adaptive chunking strategies are needed to retain key information. Documents contain precise fields such as equipment numbers, batch identifiers, and tolerance thresholds, requiring retrieval to support targeted recall by specific fields to avoid mixing irrelevant content. Additionally, this link is tied to finished product release decisions, so recalled content must be strongly correlated with the current production batch and equipment status to reduce false recall risks.
How to Configure Settings
| Config Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category of documents contains precise tolerances and judgment standards. Filtering low-relevance content prevents incorrect recall in nonconforming product release scenarios, and aligns with FastGPT v4.17.0 retrieval rules |
limit | Top 8–12 entries | The number of documents in this category grows with production batches. Limiting the number of recalled results controls context token usage while covering sufficient reference basis |
searchMode | mixedRecall | This category has both structured fields (such as equipment numbers) and text descriptions (such as inspection specifications). Hybrid retrieval balances precise matching and semantic recall |
usingReRank | Enabled | This category’s decisions are tied to quality compliance. Reranking further filters highly relevant historical defect records and inspection standards to improve recall accuracy |
datasetSearchUsingExtensionQuery | Enabled | User queries for this category often include equipment fault descriptions. Extended queries can link to historical processing records for similar faults, supplementing information not covered by the original query |
collectionIds | Bind the exclusive dataset ID for this category | Prevent recalling documents from other categories, ensuring the retrieval scope only covers content related to process quality and finished product release |
Three Common Misconfigurations
- Phenomenon: Some structured documents are imported without generating question-answer pairs, and are stored directly as original text fragments in the knowledge base. Cause: Specialized chunking and question-answer pair generation rules are not enabled for structured documents such as control plans and SPC reports, or model quota shortages cause generation tasks to fail to complete.
- Phenomenon: The retrieval interface returns the error "insufficient_quota Current group upstream load is saturated, please try again later". Cause: Retrieval concurrency limits are not configured. Synchronous retrieval requests in mass production scenarios exceed platform load thresholds, or failing to limit the number of recalled results leads to overly long contexts that increase service pressure.
- Phenomenon: Retrieval results have low relevance to user queries. For example, querying the inspection specifications for a specific model of equipment but recalling irrelevant incoming material inspection records. Cause: A reasonable
similaritythreshold is not set, or the hybrid retrieval mode ofsearchModeis not used, leading to reliance on only a single retrieval method that cannot cover structured field matching requirements.
How to Confirm Configurations Are Set Correctly
- Import a typical document (such as a finished product inspection report), check whether the retrieval can accurately recall the equipment inspection requirements in that document, and verify that the fields of the recalled content match the original document.
- Simulate batch retrieval requests, check that platform load monitoring does not show saturation errors, and confirm that concurrency current limiting configurations match the request frequency of production scenarios.
- Adjust the value range of the
similarityparameter, verify that the relevance of recall results meets the requirements of business decision-making, and ensure that false recall of irrelevant content does not occur. - Check the dataset binding status in the knowledge base, confirm that only the exclusive document collection for this category is bound, and no data sources from other categories are mixed in.
The values provided are common starting points and should be measured against the reader's own samples.