What this category of data looks like
This category of data is generated for individual transformers and high-voltage electrical products. Daily documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. Data sources cover design BOM, silicon steel sheet and insulating material procurement data, vacuum drying equipment operation data, oil test reports, and high-voltage test records from test stations. Data is updated once per product or per batch of raw materials. Core fields include insulating moisture content, partial discharge values, withstand voltage rating, nameplate parameters, and more. Full process traceability is required per individual unit.
What constraints these characteristics impose on the citation sources and traceability workflow
The requirement for full process traceability of individual products means traceability must be tied to a unique product nameplate number, to avoid mixing documents from different batches or units. Multiple heterogeneous document types, including equipment operation data, test reports, procurement records, and more, must be uniformly linked to the individual product identifier. This increases the complexity of recall matching. The high-frequency update frequency of individual product data requires traceability tools to associate the latest test and assembly records in real time, and cannot rely on static offline knowledge bases. Traceability of core audit indicators such as insulating moisture content and partial discharge values requires precise matching of original data from the corresponding test links, and cannot broadly recall irrelevant documents.
How to configure settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Matches the total length of core test reports and assembly records for individual products, to avoid truncating critical traceability data |
responseData | Enable source display | Meets compliance requirements for tracing original test data during supplier audits |
detail | Full mode | Retains original document fields and units, to avoid loss of core information such as insulating moisture content and partial discharge values |
| Maximum number of citations | Top 3 entries | Focuses on the 3 core documents for individual products, to avoid redundant information interfering with audits |
| Context window length | 1500–2000 characters | Covers key node data across the full process of individual products, to ensure contextual coherence for multi-round audits |
This configuration is based on FastGPT v4.17.0.
Three common misconfigurations
- Phenomenon: Citations of insulating oil reports not linked to the target product appear in responses. Cause: The unique product nameplate number was not used as a retrieval anchor, leading to recall of document data from cross batches.
- Phenomenon: Citation links at the bottom of responses point to incorrect documents or are inaccessible. Cause: The correct association path for
responseDatawas not configured, and temporary cached data was used as a permanent traceability source. - Phenomenon: Critical values from partial discharge test reports are truncated in responses. Cause: The value of
quoteMaxTokenwas set too low, leading to truncation of core test data that cannot fully support quality audits.
How to confirm proper configuration
- The nameplate number of an individual transformer and a query for insulating moisture content are entered, and the returned results are checked to confirm the cited documents are the vacuum drying and insulating oil reports for that specific product.
- The citation list at the bottom of the response is reviewed, and each citation is confirmed to include the corresponding document type, generation time, and core fields.
- The value of
quoteMaxTokenis adjusted, long-text partial discharge test report queries are tested, and no content truncation is confirmed. - A multi-round audit conversation is simulated, and subsequent questions are checked to still associate with the individual product data from the initial query, with no cross-batch recall.
The values given are common starting points and should be measured against the reader's own samples.