What This Category of Data Looks Like
Business data for the photovoltaic module encapsulation process comes from the cell grading system, glass adhesive film backplane batch ledger, production MES system, string welding and lamination equipment logs, and EL/IV tester output files. Daily produced documents include layout diagrams, material batch records, lamination process records, EL crack images, IV test reports, insulation pressure resistance records, and packing lists. Data is updated with one test record per module, and one batch ledger per batch of raw materials. Document fields include material model, process parameters, test values, equipment number, and more. Units cover physical quantities such as current, voltage, pressure, and time.
Constraints on Knowledge Base Retrieval and Recall
This category of data has high update frequency and fine granularity. Each module and each batch of materials generates independent records. Knowledge base retrieval must support incremental recall to avoid resource consumption from full indexing. Document types include structured ledgers, equipment logs, and visual charts. Retrieval systems must adapt field matching logic for multiple data source types. Battery cell current gear matching is a core time-consuming step. Retrieval must prioritize associating cell grading data with lamination process parameters. It is also necessary to ensure the accuracy of recall results, to avoid production issues such as non-compliant power output and lamination delamination caused by incorrect matching.
The following configurations are formulated based on FastGPT v4.17.0.
How to Set the Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Photovoltaic encapsulation process parameters require high matching accuracy to avoid interference from irrelevant process entries with low similarity in retrieval |
limit | Top 8 entries | This category requires matching multi-dimensional data (materials, processes, tests). 8 results can cover core associated information without exceeding the conventional context window |
usingReRank | Enabled | Unstructured data such as EL images and IV test reports require reranking to filter invalid matches and improve retrieval accuracy |
searchMode | embedding | Process specifications and test reports contain a large number of professional terms. Semantic retrieval can support cross-document term association matching |
datasetSearchUsingExtensionQuery | Enabled | Fuzzy questions in production scenarios can be covered by expanding queries, such as adding keywords like "lamination temperature" and "cell gear" |
collectionIds | ID set containing cell grading dataset and encapsulation process dataset | Multi-dimensional production data must be associated to ensure retrieval covers the full link of materials, processes, and tests |
Three Common Misconfigurations
- Phenomenon: Duplicate lamination process records or material batch records appear in retrieval results. Cause: No duplicate verification rule is configured for files within the dataset, or duplicate uploaded documents of the same batch are not deduplicated.
- Phenomenon: Knowledge base indexing tasks take an excessively long time to complete, and high-resolution EL images, IV test reports, and other files fail to generate indexes. Cause: The retrieval chunking strategy is not adjusted for large-volume visual documents, causing timeout blocking during the indexing process.
- Phenomenon: Retrieval results cannot associate battery cell current gears with lamination process parameters. Cause: The
collectionIdsconfiguration does not include both the cell grading dataset and the encapsulation process dataset, causing cross-data source association retrieval to fail.
How to Verify the Configuration Is Properly Set
- Submit a test question containing professional terms, such as "How to handle non-compliant string welding spacing". Check whether the returned results include corresponding lamination process records.
- Upload a duplicate material batch record. Check whether the dataset automatically completes deduplication verification.
- Trigger an indexing task. After the task is completed, check whether the index status of documents such as EL images and IV test reports is normal.
- After enabling the
usingReRankconfiguration, submit a test question with a vague description, such as "How to retest a module with EL cracks". Check whether the reranked results prioritize displaying process specifications related to EL images.
The values provided are common starting points and should be measured against the reader's own samples.