What This Category of Data Looks Like
Data sources for this category include MES systems, online inspection equipment, laboratory test results, process parameter logs, customer quality standards, and historical defect libraries. Data updates run per production batch and per shift. Some real-time collected process and test data sync continuously. Daily documents include control plans, patrol inspection records, SPC reports, finished product inspection reports, nonconforming product review records, release orders, and quality monthly reports. Data fields include batch identification, inspection item name, defect type, process parameter value, inspection result, release status, and more. Corresponding units are determined by inspection items and process requirements.
Constraints for Knowledge Base Retrieval and Recall
This category’s data characteristics create multiple constraints for retrieval and recall. Multi-source, scattered, and frequently updated data requires support for incremental sync and permission isolation. This ensures retrieval scope covers the latest batch and defect data. Diverse document formats range from structured inspection reports to non-standardized review records. Adaptive text splitting rules are needed to avoid losing critical associated information after splitting. Fields adjust dynamically based on inspection items. Retrieval must accurately match core fields such as batch identification and defect type. This filters out irrelevant results. Retrieval results must also link to the permission scope of finished product release. This prevents unauthorized access to restricted quality documents.
Configuration Settings
All configurations below are based on FastGPT v4.17.0. They are adapted to retrieval needs for process quality and finished product release scenarios.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filters low-relevance quality documents. It retains sufficient defect cases and inspection standard content to meet this category’s semantic matching needs. |
limit | Top 10 entries | Controls the context length of retrieval results. It avoids exceeding model input limits and covers defect records and inspection items across multiple batches. |
searchMode | embedding | Adapts to this category’s core semantic matching needs. It accurately matches quality data associated with defect types, process parameters, and other relevant fields. |
usingReRank | Enabled | Corrects semantic similarity deviations across multi-source data. It improves matching accuracy between results and current traceability queries. |
datasetSearchUsingExtensionQuery | Enabled | Automatically adds pre-filter conditions such as batch and defect type based on user traceability queries. This narrows the retrieval scope. |
collectionIds | Configured in groups by business department and document type | Precise limitation of retrieval scope. Only includes inspection records, review records, release orders, and other documents related to process quality and finished product release. |
Three Common Misconfigurations
- Phenomenon: Retrieval results include unauthorized release order documents. Inspection records with no valid content are returned. Cause:
collectionIdsis not configured to limit retrieval dataset groups. Documents with no valid inspection results are not filtered. - Phenomenon: Responses exceed the knowledge base scope. Uncollected process standards or customer requirements are included. Cause: Retrieval scope is not strictly limited. The configuration restricting responses to only knowledge base content is not enabled.
- Phenomenon: Retrieval result count is far lower than expected. Historical defect data for the same batch is not covered. Cause:
similaritythreshold is set too high. Some semantically relevant historical defect records with slightly lower similarity are filtered out.
How to Verify Proper Configuration
- A test set of control plans and defect records for a single batch is uploaded. A retrieval including specified batch keywords is initiated. Returned results are confirmed to only include relevant documents for that batch.
- A retrieval request with no matching content is initiated. A preset custom response is confirmed to be returned. No information outside the knowledge base is included.
- Dataset group configurations are reviewed. Retrieval scope is confirmed to only include document sets related to process quality and finished product release. Unrelated incoming material inspection documents are excluded.
- Retrieval logs are checked.
usingReRankanddatasetSearchUsingExtensionQueryare confirmed to be active. Retrieval results are sorted by matching degree from highest to lowest.
The values provided here are common starting points, and should be evaluated against the user’s own samples.