What This Category of Data Looks Like
This category of data originates from process, equipment, quality control, and production departments of crystalline silicon cell manufacturers. Daily documents include process cards, diffusion and coating parameters, printing and sintering records, EL detection images, IV grading data, and yield reports. Data sources include silicon wafer incoming materials, MES systems, diffusion furnaces, coating equipment, printing presses, EL and IV testing devices, and laboratory data. Updates occur synchronously with each production batch and each piece of tested material. Some reports are aggregated on a daily basis. Document fields contain specific process parameters, test result values, electrical performance grading identifiers, and yield metrics. These cover professional parameter units such as temperature, time, and current efficiency.
What Constraints Do These Characteristics Impose on Knowledge Base Retrieval and Recall
The high-frequency updates, mixed multiple document types, and strong internal correlation of this category of data create multiple constraints for knowledge base retrieval and recall. High-frequency updates require the retrieval pipeline to support near-real-time synchronization, to prevent outdated data from interfering with process judgment. Multiple document types—including structured parameters and graphic-text detection reports—require adaptive multimodal indexing rules. These rules must support unified retrieval logic for numerical values, images, and text. Professional fields and units require the retrieval model to accurately match industry terminology, to reduce ambiguous matches. Strongly correlated data grouped by batch and equipment dimensions require support for filtering recall results by production node. This ensures retrieved content aligns with the current production scenario.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale for This Setting |
|---|---|---|
similarity | 0.75–0.85 | Photovoltaic process parameters require high matching accuracy to avoid interference from low-relevance general process content in professional retrieval |
limit | Top 8–12 items | The number of process-related documents per batch is moderate. Too many will exceed the context window, while too few will miss critical parameters |
searchMode | mixedRecall | Balances semantic matching and precise recall of professional keywords, adapting to the mixed retrieval needs of structured parameters and unstructured reports |
usingReRank | Enabled | When multiple types of documents are mixed, re-ranking can improve the sorting priority of highly relevant results and reduce the chance of professional parameters being overshadowed by non-critical content |
datasetSearchUsingExtensionQuery | Enabled | Extended queries can supplement process-related synonyms and industry terminology to cover the question-asking habits of different engineers |
collectionIds | Bind three dedicated collections for process, equipment, and quality control | Dividing collections by business department can accurately filter irrelevant documents and narrow the retrieval scope to improve efficiency |
Three Common Configuration Errors
- Phenomenon: A large amount of general process content unrelated to photovoltaic cells is mixed into retrieval results, and the number of results far exceeds the set
limit. Cause: The dedicatedcollectionIdsare not bound by business department, so the retrieval scope covers the entire platform knowledge base, and irrelevant documents are not filtered. - Phenomenon: Professional parameters such as "diffusion sheet resistance" cannot be retrieved, and the recall results are empty. Cause:
datasetSearchUsingExtensionQueryis not enabled, and industry synonym extensions are not added, so professional question phrasing from engineers cannot be matched. - Phenomenon: The sorting of retrieval results is chaotic, with highly relevant process cards ranked behind low-relevance detection reports. Cause:
usingReRankis not enabled. The basic recall sorting only relies on vector similarity, and no secondary sorting optimization is performed for multiple types of documents.
How to Verify a Successful Configuration
- Retrieve specified process parameters, and check whether the returned results only include dedicated documents for photovoltaic cell manufacturing, with no cross-category content.
- Enter professional terminology such as "EL hidden crack detection", and confirm that the retrieval results include corresponding detection reports and process descriptions, with no obvious omissions.
- Adjust the value range of the
similarityparameter, observe changes in the relevance of retrieval results, and ensure that they meet the accuracy requirements of the business scenario. - View the retrieval log to confirm that the number of recall results conforms to the
limitsetting, with no abnormal overage or underage.
The values provided here are common starting points and should be validated against local operational samples.