This page is compatible with FastGPT v4.17.0.
What This Category of Data Entails
Data sources for this category include material grades, customer technical specifications, temperature control instruments, furnace temperature uniformity test data, hardness tester and metallographic equipment data, and MES system data. Daily documentation covers heat treatment process cards, charging charts, furnace temperature curves marked with ℃ and minute units, hardness and metallographic reports containing HV/HRC hardness values and metallographic structure ratings, equipment inspection forms with equipment numbers and inspection times, and exception handling sheets. Data is updated per heat and per batch. Documents associated with a single heat have high relevance and a fixed number of entries.
Constraints on Citation Sources and Traceability
First, data is updated per heat. Traceability must accurately bind all associated documents for a single heat to avoid mixing data across heats. Second, documents contain numerical and rating fields with clear units. When citing, field units and parameters must be matched to prevent mixed hardness value units or mismatched furnace temperature and time values. Third, data sources are scattered across devices and systems. Unique identifiers for multi-source data, such as heat numbers and equipment numbers, must be integrated to ensure traceability back to the original collection device. Fourth, the number of associated documents per single heat is fixed. Context recall scope must be limited to avoid redundant data interfering with review results.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Covers the total character count of core documents for a single heat treatment, and can accommodate core fields of process cards, furnace temperature curves, and hardness reports |
responseData | Enabled | Required to return unique identifiers such as heat numbers, material grades, and equipment numbers to support cross-document traceability |
detail | Enabled | Required to display the document type, generation time, and data source equipment corresponding to each citation |
| Single recall limit | Top 6 entries | Effective documents associated with a single heat do not exceed 6 categories, avoiding context overload |
| Traceability matching threshold | Calibrated via actual testing | Matches the field similarity requirements for heat treatment documents, and adapts to the association logic of multi-source data |
| Source verification switch | Enabled | Verifies the consistency of heat numbers and equipment numbers in cited data to prevent incorrect cross-heat associations |
Three Common Misconfigurations
- Phenomenon: The response body does not include matched local knowledge base document content, but a corresponding document entry exists in the citation list. Cause: The
detailconfiguration is not enabled, and core fields of knowledge base documents are not injected into the response context. - Phenomenon: When the API is called to reference a private knowledge base, a no matching results prompt is returned. Cause: No binding index is configured between the knowledge base heat number and the process card, so the heat identification parameter in the submitted query cannot be matched.
- Phenomenon: Cross-heat incorrect associations appear in cross-review results, such as binding the furnace temperature curve of heat A with the process card of heat B. Cause: The source verification switch is not enabled, and irrelevant documents are incorrectly matched due to keyword-only recall.
How to Verify Correct Configuration
- A single-heat heat treatment process card, furnace temperature curve, and hardness report are uploaded. A cross-review request is initiated, and whether the response body includes the core fields of the three document types is verified.
- The API is called with a query containing a clear heat number, and whether the citation list in the returned results includes documents for the corresponding heat, and the body matches the relevant content, is checked.
- System operation logs are reviewed to confirm that the number of context recalls per request does not exceed the set single recall limit, and no timeout errors occur.
- Returned citation source information is manually verified to confirm that the document type and data source equipment number match the actually uploaded documents.
The values provided are common starting points and should be measured against the reader's own samples.