What This Category’s Data Looks Like
The data for photovoltaic intelligent due diligence reports primarily comes from power station operation and maintenance logs, component factory inspection documents, grid connection reporting data, and regional meteorological irradiation datasets. The data update schedule covers real-time operation logs, daily power generation statistics, monthly grid connection reports, and quarterly batch inspection reports. Individual documents are mostly structured tables or long-text and image reports, containing fields such as unique power station identifier, component model, irradiation intensity (unit: kWh/㎡), power generation capacity (unit: kW), fault code, and maintenance cycle. Some archived reports also include photo attachments from on-site inspections.
What Constraints These Characteristics Impose on Citation Sources and Traceability
Photovoltaic category data is multi-source, heterogeneous, and fields have clear units. This requires precise matching of field names and units during citation traceability to avoid confusion between identically named fields from different power stations. Data formats include structured logs, text-image reports, and attachment files. During recall, it is necessary to distinguish document types and associate corresponding original storage paths and generation times. The high-frequency update feature of real-time operation logs requires that data source synchronization configurations match the update schedule to ensure cited content is up-to-date and valid. For components of the same model, there are multiple batch inspection reports. Traceability requires additional matching of production batch identifiers to prevent cited content from mismatching the actual parameters of the corresponding power station.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall Count | Top 8-12 entries | The cited content of photovoltaic due diligence reports is mostly long fragments associated with multiple fields. Excessive recall will lead to redundant context, while insufficient recall will fail to cover core operation and maintenance and inspection data of power stations |
Similarity Threshold | 0.72-0.85 | Photovoltaic data contains a large number of semantically similar fields, such as "irradiation intensity" and "sunlight exposure". A threshold that is too low will introduce irrelevant recalls, while a threshold that is too high will miss valid matching content |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Photovoltaic long-text and image inspection reports contain a large number of tables and image analyses, requiring sufficient time for text splitting and metadata extraction |
Citation Deduplication Switch | Enabled | Operation logs of the same power station may be uploaded to the knowledge base repeatedly. Enabling this switch avoids multiple citations of the same document fragment and ensures unique traceability information |
Traceability Information Retention Fields | Power Station ID, Component Model, Generation Time, Original File Name | Traceability of photovoltaic due diligence reports requires clear correspondence to power stations and document batches. These fields allow quick location of original data sources |
Reranked Return Count | Top 5-7 entries | Reranking optimizes the relevance ranking of recalled content, matches the multi-field association feature of photovoltaic data, and reduces invalid citations |
The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Configuration Mistakes
- Phenomenon: After 100 photovoltaic-related files are uploaded to the same knowledge base, about half show training exceptions. Re-uploading still results in duplicate traceability during citation. Cause: Some files have incorrect metadata formats (such as the Power Station ID field), leading to failure to correctly associate traceability information during parsing, and citation deduplication is not enabled, so duplicate data is not filtered out.
- Phenomenon: After setting the citation upper limit for semantic retrieval, the returned cited content contains a large amount of irrelevant data from non-photovoltaic power stations. Cause: The similarity threshold is not adjusted based on the semantic characteristics of photovoltaic data, and the recall count is set too high, introducing a large number of low-relevance contents.
- Phenomenon: When using
{{id}}in the cited content template, the displayed value is the ID of the knowledge base collection, not the ID of a single cited data entry. Cause: The traceability information retention fields are not correctly configured, and the unique identifier of a single data entry is not extracted, causing the template to fail to correctly obtain the ID of the cited data.
How to Confirm Proper Configuration
- Upload a single photovoltaic inspection report, trigger knowledge base parsing, and check if the parsed metadata contains the preset traceability retention fields to confirm that the fields are complete.
- Initiate a due diligence query for a photovoltaic power station, check if the returned cited content is attached with the correct original file name and generation time to confirm that the traceability information is correctly associated.
- Upload two duplicate photovoltaic operation logs, check if the citation list only retains one valid traceability record to confirm that the deduplication function is effective.
- Adjust the similarity threshold and initiate a query, compare the number of recalled contents before and after to confirm that the impact of the threshold configuration on recall results meets expectations.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.