This content is based on FastGPT v4.17.0, and is adapted for the inspection and process record review scenario for in-process quality and final product release in the manufacturing industry.
What the Data for This Category Looks Like
The data for this category comes from manufacturing enterprise MES systems, online inspection equipment, laboratory systems, process parameter libraries, customer standard libraries, and historical defect libraries. Daily documents include control plans, patrol inspection records, SPC reports, final product inspection reports, non-conforming product review forms, release forms, and quality monthly reports. Each document contains fields such as inspection items, tolerance values, inspection frequencies, defect types, and release thresholds. Units cover length, weight, defect rate, and others. Teams collect data synchronously with each production batch and each shift’s patrol inspection. Some real-time monitored process parameters receive continuous updates.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Scattered data sources and uneven update frequencies require retrieval systems to match documents across multiple heterogeneous formats, and support unified recall logic across data sources. Documents include precise fields such as professional tolerances and defect thresholds. Retrieval results must strictly match business rules associated with these fields, and avoid generalized recall. High-frequency updates per batch and per shift require the retrieval link to support near-real-time incremental updates, and prevent recall of stale data. The strong association between historical defect libraries and release forms requires retrieval to recall related documents according to business scenarios. Returning isolated single records does not meet business requirements.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches precise business fields such as tolerances and thresholds, and ensures relevance between recall results and inspection standards |
limit | Top 6 entries | Combined with the associated attributes of non-conforming review and release forms, a single recall sufficiently covers multi-dimensional inspection basis |
searchMode | mixedRecall | Supports hybrid retrieval of text and structured fields, and adapts to field matching for multiple types of documents |
usingReRank | Enabled | Re-ranks recall results by business relevance, and prevents results with precise field matching but low business relevance from appearing first |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands inspection-related terms based on user queries. For example, expands "dimension deviation" to expressions corresponding to tolerance ranges |
collectionIds | Specify dedicated collections for quality documents | Only retrieves category-specific documents such as control plans and inspection reports, and avoids mixing in irrelevant data |
Three Common Misconfigurations
- Retrieval results are empty or fail to recall specified documents. Check logs to confirm that similarity matching did not hit core fields. The cause is that
datasetSearchUsingExtensionQueryis not enabled, and professional terms are not expanded for matching. This causes user queries to fail to align with document fields. - A single API call returns a maximum of 6 entries, and the quantity cannot be adjusted higher. The cause is failing to confirm the correct modification path for the retrieval upper limit parameter in the private deployment configuration file. The default limit does not match business requirements.
- New knowledge bases can recall documents normally, but original knowledge bases fail. The cause is that the
collectionIdsconfiguration of the original knowledge base did not correctly associate the current business collection, or core fields such as inspection items and tolerance values were not retained during document chunking.
How to Confirm the Configuration Is Correct
- Submit a query containing inspection items and tolerance ranges, and verify if returned results include category-specific documents such as control plans and final product inspection reports.
- View the order of retrieval results. Confirm that after
usingReRankis enabled, results with high business relevance appear first. - Adjust query keywords, and verify whether
datasetSearchUsingExtensionQueryautomatically expands relevant professional terms and returns matching historical defect library documents. - Call the API interface, and confirm that the number of returned results matches the configured
limitvalue, and is not affected by the default limit.
The values provided here are common starting points, and should be evaluated against the samples used in the actual deployment.