What This Category of Data Looks Like
Data for this category originates from material grade standards, customer technical requirements, temperature control instruments, furnace temperature uniformity test results, hardness tester and metallographic equipment test data, and manufacturing execution system (MES). Daily generated documents include heat treatment process cards, furnace loading layout drawings, furnace temperature curve records, hardness and metallographic test reports, equipment inspection sheets, and exception handling documents. Data is updated per furnace load and per batch. Fields include heating temperature, holding duration, cooling rate, hardness value, and metallographic structure grade. Common units are degrees Celsius, hours, HRC, and metallographic grades.
Constraints Imposed on Knowledge Base Retrieval and Recall
This category of data is updated frequently per furnace load and per batch. Knowledge base retrieval must limit the search scope to documents associated with the current production batch to avoid interference from cross-batch data. Multiple types of documents (process cards, furnace temperature curves, test reports) are stored together, so cross-document associated recall must be supported. For example, combine furnace loading layout drawings and furnace temperature curves to verify process compliance. Professional fields have dedicated units, so search must retain unit matching precision to prevent non-precise matches. Frequently updated equipment inspection and exception handling data require incremental indexing logic to avoid delays from full index rebuilding.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for heat treatment terminology and units, prevents irrelevant documents with low matching scores from being recalled |
limit | Top 8 entries | The number of heat treatment-related documents per batch is limited, controlling the number of returned entries to avoid context overload |
searchMode | mixedRecall | Combines keyword and semantic retrieval, adapts to the retrieval needs of mixed-type documents such as text descriptions in process cards and numerical records in furnace temperature curves |
usingReRank | Enabled | Reranks initial recall results, prioritizes matches for core process parameters such as furnace loading position and furnace temperature curves |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms related to processes, such as associating "quenching" with professional terms like "hardening" and "martensitic transformation" |
collectionIds | Bound to workshop/batch | Limits the search scope to knowledge base collections associated with the current production batch, avoids interference from cross-batch data |
Three Common Misconfigurations
- Symptom: After uploading a standard heat treatment process card, it is displayed in 8 segments, then changes to 13 segments after refreshing, with duplicate index entries. Cause: The deduplication check logic for incremental indexing is not enabled, and the chunk overlap parameter is set incorrectly, causing the same segment of content to be split and indexed multiple times.
- Symptom: Retrieval results only list citation entries at the end of the response, and cannot annotate matching furnace temperature curves or hardness reports at corresponding positions in the main text. Cause: The context anchoring function for knowledge base citations is not configured, only a global citation list is retained.
- Symptom: Retrieval results do not include furnace temperature curve data for the current batch, leading to failure of process compliance verification. Cause: The
collectionIdsparameter is not correctly bound, the search scope is not limited to the knowledge base collection of the current production workshop, and documents from other batches are mixed in.
How to Verify Proper Configuration
- Upload a standard heat treatment process card, check that the number of chunks is stable with no duplicate split entries, verify that the deduplication logic is active.
- Enter a query containing furnace temperature and holding duration, check that retrieval results prioritize returning the corresponding furnace temperature curves and process cards, verify that the reranking and query expansion functions are active.
- Bind a knowledge base collection for a specified workshop, enter a cross-batch query, check that retrieval results only include documents within the currently bound collection, verify that the scope restriction is active.
- View the citation list in retrieval results, confirm that each citation corresponds to the uploaded original document with no redundant entries, verify that the index configuration is correct.
The values provided are common starting points and should be evaluated against the reader's own samples.