What the Data for This Category Looks Like
Data from the photovoltaic module encapsulation process comes from the cell sorting system, glass adhesive film backplane batch ledger, workshop MES system, string welding and lamination production equipment, and EL/IV testers. Daily retained documents include layout drawings, material batch records, lamination process records, EL defect images, IV test reports, insulation voltage resistance records, and finished product packing lists. One set of exclusive records is created per module. Raw material batch ledgers are updated synchronously for each batch. Each single record includes process parameters, test results, and associated material information for the corresponding process step.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
The data characteristics of this category impose multiple constraints on the retrieval and recall link. Multiple types of documents coexist, including static process drawings, numeric test reports, and real-time process parameters. Precise cross-type content matching must be supported. Data updates frequently, and single records are tightly associated. Vector database synchronous update efficiency must be ensured to avoid retrieval results being disconnected from current production data. Single records contain multi-dimensional information: associated materials, processes, and test results. Retrieval must prioritize content strongly bound to current process parameters to reduce invalid matches. Test result interpretation relies on precise parameter matching. Recall result redundancy must be limited to improve audit efficiency.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Precise matching of process parameters and test results is required to avoid irrelevant low-similarity content being included |
limit | top 8 entries | Single module records have many associated information items. Too many results exceed the context window. Too few fail to cover associated process content |
searchMode | mixedRecall | Documents include drawings, numeric reports, and text records. Hybrid mode balances image-text matching and parameter keyword recall |
usingReRank | enabled | Multiple types of documents coexist. Reranking filters low-relevance recall results and improves final retrieval accuracy |
datasetSearchUsingExtensionQuery | enabled | Process parameters are tightly associated with material batches. Extended queries supplement associated batch information and improve matching coverage |
collectionIds | bind three dedicated knowledge base collections for string welding, lamination, and testing | Adapts to FastGPT v4.17.0 knowledge base collection binding logic. Dividing data by production process avoids irrelevant cross-process content interference |
Three Common Misconfigurations
- Phenomenon: Retrieval results include expired batch records from other workshops. Cause: No vector database auto-refresh task is configured. Old batch data is not cleaned timely and mixed with current production data.
- Phenomenon: Returned EL images do not match corresponding IV test reports. Cause: Knowledge base collections are not divided by string welding, lamination, and testing processes. Irrelevant cross-process content is incorrectly recalled.
- Phenomenon: Retrieval results include unauthorized externally published documents. Cause: The internet retrieval switch is not disabled, and the system is not restricted to specified internal knowledge base collections. External content is included in the recall scope.
How to Confirm the Configuration Is Correct
- Upload a test lamination process record, enter parameter keywords for the corresponding process step, and verify the retrieval result prioritizes returning this record.
- Check knowledge base permission settings to confirm only authorized personnel can access private knowledge base collections, preventing external content from being included.
- Run a batch retrieval test, count retrieval time and recall result count, and adjust relevant configurations to meet business requirement thresholds.
- Verify the vector database refresh task is executing normally. Confirm expired data is cleaned regularly to avoid old records interfering with retrieval.
The values listed above are common starting points and should be evaluated against the reader's own samples.