What the data for this category looks like
Data for power grid equipment intelligent due diligence reports comes from equipment factory inspection documents, on-site inspection ledgers, bidding technical specifications, infrared temperature measurement inspection images, and similar sources. Update frequency varies widely by document type:
- Factory documents are updated once upon project delivery.
- Inspection ledgers are updated monthly or quarterly.
- Operation ledgers synchronize real-time operating data.
Document structures primarily include structured parameter tables, long-text operation records, and technical images with embedded text. Core fields include rated voltage (unit: kV), rated current (unit: A), insulation resistance (unit: MΩ), equipment operating duration (unit: hours), batch number, model code, and other relevant fields.
Constraints imposed on document parsing and chunking
High proportions of structured parameter tables require accurate identification of table structures and cell correspondences during parsing, to avoid losing associations between parameters after splitting. Technical images with embedded text require extraction of both in-image values and annotations, rather than only parsing plain text content. Long-text operation records may span multiple pages. Chunking must retain contextual connections to prevent records from the same inspection cycle from being split across different chunks. Fields include clear physical units. Parsing must bind units to their corresponding parameters to avoid separating values and units. When multiple source documents are uploaded together, chunks must be categorized by document type to prevent mixing device data from different batches.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
parse_table | true | Power grid equipment due diligence reports contain many parameter tables. Retaining structured content ensures full parameter completeness |
image_ocr_enable | true | Documents include infrared temperature measurement images, wiring diagrams, and other images with embedded text. Extraction of device parameters from image annotations is required |
chunk_size | 800–1200 characters | Balances device parameter completeness and contextual relevance, avoiding excessive splitting of long operation records |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Single device documents often include multiple pages of inspection records and images. Sufficient time is needed to complete full parsing |
table_chunk_split_mode | keep_table_whole | Prevents device parameter tables from being split, ensuring complete correspondence of all parameters within a single chunk |
recall_threshold | 0.75 | Filters low-relevance chunk content, improving the accuracy of device parameter retrieval |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Parsed device parameters are output as plain text, and cell alignment and field correspondence are not retained. Cause: The
parse_tableconfiguration is not enabled, and only plain text content from the document is extracted. - After calling the knowledge base creation API, real-time statuses for parsing in progress, ready, and failed cannot be retrieved. Cause: The
parse_statusfield in the returned results is not monitored, or a webhook is not configured to receive parsing status push notifications. - An empty value is returned when attempting to parse the Set-Cookie field. Cause: Network outbound permissions are not configured to allow access to the target interface, or parsing rules do not match the Set-Cookie field in the response header.
How to Verify Correct Configuration
- Upload a test document containing device parameter tables and infrared temperature measurement images, and confirm that the parsed content retains the structured layout of the table and the text values within the image.
- Call the knowledge base creation API, check that the returned results include the
parse_statusfield, and verify that subsequent status updates align with actual parsing progress. - Review the chunked content list, and confirm that a single device parameter table exists as an independent chunk, without being split into multiple segments.
- Initiate a device parameter retrieval request, and check that the returned chunked content matches the parameters in the uploaded document exactly.
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.