What Data for This Category Looks Like
This category typically includes documents such as process cards, diffusion and coating parameter logs, printing and sintering records, EL detection images, IV grading data, and yield reports. Data sources include silicon wafer incoming material ledgers, production MES systems, operating data from diffusion furnaces, coating equipment, and printing presses, EL and IV test results, and laboratory quality inspection records. Update cycles cover each production batch, single-wafer detection nodes, and daily summary processes. Fields include batch numbers, equipment operating parameters, detection image identifiers, electrical performance values, appearance defect markers, and other relevant fields.
Constraints Imposed by These Characteristics on the Citation Source and Traceability Link
Data for this category includes both single-wafer level detection details and batch level summary information. Citation traceability must accurately match production batches or single-wafer numbers to avoid cross-batch confusion. Multi-source heterogeneous data, including structured parameters and unstructured images, requires the retrieval system to associate documents and test results in different formats. Frequently updated production and detection data requires the traceability link to quickly retrieve the latest process and quality inspection records. Strong inter-process correlation requires cited sources to cover corresponding data from continuous processes including texturing, diffusion, passivation, printing and sintering, to ensure that quality problems can be fully traced back to the corresponding process window.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum citation token count | Set quoteMaxToken to 800–1200 tokens | Photovoltaic cell process parameters and detection records have long individual entry lengths, ensuring a single citation can cover a complete parameter set or detection result |
| Response data includes | Select "Source document metadata + snippet" for responseData | Citation source details including batch number, equipment ID, and detection time must be recorded to support accurate traceability positioning |
| Retrieval detail level | Set detail to "Full snippet" | Unstructured EL detection images and IV data require complete snippet content to be cited, enabling subsequent verification of the correspondence between defects and electrical performance |
| Number of recalled citations per request | Top 5 entries | Quality issues for this category are typically associated with multi-process data, and around 5 entries can cover citation sources for the main associated links |
| Similarity threshold | 0.75–0.85 | Low-correlation general parameters must be filtered out to accurately match process window data for specific batches |
Three Common Misconfigurations
- Phenomenon: When a knowledge base ID or static citation variable is entered directly, output results are consistent. After binding application variables, a system error is triggered. Cause: The mapping relationship between application variables and knowledge base metadata is not correctly configured, or the variable call format does not comply with FastGPT syntax rules.
- Phenomenon: Mixed retrieval returns citations that include irrelevant component packaging documents, failing to filter target photovoltaic cell manufacturing data. Cause: No document type filtering rules for retrieval are set, and the data source is not limited to process and quality inspection documents from the photovoltaic cell manufacturing link.
- Phenomenon: Recalled citation snippet length exceeds expectations, leading to context overflow. Cause: No reasonable upper limit for
quoteMaxTokenis set, or the number of single recall entries is configured too high.
How to Confirm Proper Configuration
- Initiate a quality audit query that includes a specific batch number, and check if returned citation sources include the process card and detection records for the corresponding batch.
- Review cited metadata fields to confirm inclusion of key traceable information such as batch number, equipment ID, and detection time.
- Adjust citation-related configuration items to verify that the number of citations and snippet length of output results match expected settings.
- Test the query process with bound application variables to confirm that variables can be correctly mapped to corresponding data in the knowledge base without triggering errors.
The values provided are common starting points and should be measured against the reader's own samples.