What the data for this category looks like
Heat treatment scenario data comes from material grade standards, customer technical specification documents, real-time temperature control instrument data, furnace temperature uniformity test reports, test results from hardness testers and metallographic equipment, and workshop MES system logs. Data is updated per furnace run or production batch. Single batch data includes documents such as heat treatment process cards, furnace loading layout drawings, furnace temperature-time curves, hardness and metallographic analysis reports, equipment inspection forms, and exception handling sheets. Document fields include furnace temperature (unit: ℃), holding time (unit: hours), material grade, hardness value (unit: HV or HRC), metallographic structure grade, furnace loading position number, and others. Some documents include visual charts and tabular data.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Since data updates per furnace run with frequent incremental changes, the retrieval link must support fast incremental recall. Full scans cause performance loss and must be avoided. Multiple document types coexist, including plain text, tables, and visual charts. Retrieval and recall must support multimodal content. Professional fields such as metallographic structure grade and furnace temperature curve parameters require high matching accuracy. This prevents irrelevant non-heat treatment documents from being recalled. A single furnace run has a limited number of associated documents, but individual documents have high content density. The context window size of recall results must be controlled to stay within model processing limits.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Heat treatment data has high professionality. A high similarity threshold is needed to filter irrelevant results and prevent documents from non-target processes from being recalled |
limit | Top 8–12 entries | The number of documents related to a single furnace run is typically 5-10. Recalling too many documents will exceed the large model context window and reduce answer accuracy |
searchMode | fullTextRecall | The scenario includes non-plain-text content such as furnace temperature curves and furnace loading drawings. Full-text search can cover professional documents with mixed text and images |
usingReRank | Enabled | There are many professional terms. The re-ranking link can optimize the relevance ranking of retrieval results and improve the matching accuracy of exception reports and process documents |
datasetSearchUsingExtensionQuery | Enabled | Terms such as material grades and customer technical specifications have synonyms or abbreviations. Extended queries can improve recall rates for cross-term matching |
collectionIds | Bind dedicated knowledge base collections for heat treatment | Isolate knowledge base data from other production categories to avoid recalling irrelevant mechanical processing or surface treatment documents |
Note: This configuration is based on FastGPT v4.17.0
Three common mistakes
- Phenomenon: When retrieving Excel files, the returned matching data is less than the source file. For example, entries marked for overseas use cannot be recalled. Cause:
datasetSearchUsingExtensionQueryis not enabled, and complete cross-row and cross-column data association is not retained during chunking, resulting in failure of professional field matching. - Phenomenon: An
Error: write EPROTprompt is displayed when calling the HTTPS interface of the self-owned system. Cause: Correct interface timeout parameters are not configured, or HTTPS two-way authentication is not handled, making it impossible to stably establish a connection with the knowledge base. - Phenomenon: Process card images uploaded to the knowledge base cannot be displayed in responses, or duplicate files are not automatically cleaned up. Cause: Multimodal retrieval configuration is not enabled, or the duplicate file verification function is not enabled, a dedicated collection is not bound, or a deduplication rule based on file content hash values is not set.
How to confirm the configuration is correct
- Enter typical process parameters for a single furnace run, perform a retrieval test, and check whether the recalled document types and fields match current business requirements.
- Upload test Excel files and image files, verify that the retrieval results can correctly associate corresponding attachment content, and confirm the validity of multimodal and tabular data recall.
- Call the interface to verify that the number of returned results matches the source file, adjust the
similarityandlimitparameters to meet business requirements. - Check the interface call logs to confirm there are no errors such as
Error: write EPROT, and that the timeout parameters match the current retrieval load.
The provided values are common starting points and should be evaluated against the reader’s own operational samples.