What this category of data looks like
Data for this category comes from material grades, customer technical specifications, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers and metallographic equipment, and MES systems. Daily documents include heat treatment process cards, charging charts, furnace temperature curves, hardness and metallographic reports, equipment inspection forms, and exception handling sheets. Data is updated per furnace load or batch. Single furnace load documents include parameters such as heating temperature (degrees Celsius), holding time (hours), cooling rate, as well as HV or HRC values from hardness testing. Some documents include visual charts and equipment status records.
What constraints these characteristics impose on the "knowledge base retrieval and recall" link
The high-frequency per-furnace update requirement means retrieval and recall must support fast incremental synchronization, to avoid data lag that affects audit efficiency. Multiple document types (including structured parameters, unstructured text, and visual charts) require retrieval to cover both keyword matching and semantic understanding, to avoid missing key process details. The high matching accuracy requirement for professional parameters means low-relevance results must be filtered, to prevent incorrect process parameters or testing data from being used for audit judgments. In addition, single-batch data has a moderate volume but strong professionalism, so the number and length of recall results must be controlled, to avoid exceeding the context processing window.
How to configure settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Heat treatment processes and testing parameters have high requirements for professional matching accuracy. Filtering low-relevance results ensures the accuracy of audit basis |
limit | Top 8 entries | The total character count of documents related to a single furnace load is moderate. Excessive recall content will exceed the context window and reduce large model processing efficiency |
usingReRank | Enabled | Semantic matching varies across different document types. Reranking can improve the ranking priority of high-relevance results and optimize the quality of audit references |
searchMode | mixedRecall | Covers both structured parameter retrieval and unstructured text semantic matching, adapting to the diverse document types of this category |
datasetSearchUsingExtensionQuery | Enabled | User audit questions often involve extended matching of process parameters. Extended queries can cover process requirements associated with material grades |
collectionIds | Bind dedicated heat treatment knowledge base collections | Isolate data from other categories to avoid irrelevant documents interfering with audit retrieval results |
Three common mistakes
- Phenomenon: Non-heat treatment category documents appear in retrieval results, and some professional fields are empty. Cause: The
collectionIdsbinding for the dedicated knowledge base collection was not configured, leading to cross-collection recall of irrelevant data. - Phenomenon: Locally uploaded image links in retrieval results show as expired and cannot be viewed normally. Cause: The storage policy after local file upload was not adjusted to permanent storage, and a temporary expired storage configuration was used by mistake.
- Phenomenon: Large model responses exceed preset word limits, or cannot fully cover the process details required for audits. Cause: The
limitparameter was not adjusted based on the total character count of single furnace load documents, leading to too much or too little recalled context content.
How to confirm the configuration is correct
- Upload a heat treatment furnace temperature curve document, initiate a query that includes target furnace temperature parameters, and verify that the returned results include relevant fragments of this document.
- View the knowledge base configuration page, confirm that
collectionIdshas specified the dedicated heat treatment-related dataset, and no irrelevant collections are bound. - Initiate a query that includes an exception handling scenario, and verify that after
usingReRankis enabled, high-relevance exception handling sheet results appear at the top of the list. - Upload a heat treatment charging chart, initiate a query that includes charging position requirements, and confirm that associated text descriptions of this image can be recalled.
The values provided here are common starting points and should be evaluated against the reader's own samples.