What Data for This Category Looks Like
Data for this category comes from material grade standards, customer technical requirements, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers, metallographic equipment, and production MES systems. Daily documents include heat treatment process cards, furnace loading diagrams, furnace temperature curves, hardness and metallographic reports, equipment inspection forms, and exception handling records. Fields include furnace temperature values (unit: ℃), furnace loading station numbers, material grades, hardness values (unit: HV or HRC), and more. Data updates occur per furnace run and per batch. Single batch data is linked to complete production process records for the corresponding furnace run.
Constraints on Knowledge Base Retrieval and Recall
This category has three core characteristics: high-frequency small-batch updates, multiple heterogeneous document types, and quantified fields with units. These characteristics create clear constraints for retrieval and recall. Real-time data for each furnace run must support fast matching of corresponding process parameters to avoid excessive retrieval delays. Heterogeneous documents including curves, tables, and text must support retrieval rules tailored to different fields. Quantified indicators with units such as furnace temperature and hardness must support precise matching linked to their units. Exception analysis requires linking to historical exception handling records. Retrieved results must cover full-process linked data to avoid fragmented retrieval outcomes.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches quantified indicators and text content for heat treatment processes and exception records. This range filters irrelevant general process documents, retains highly relevant furnace temperature and hardness matching results, and aligns with FastGPT v4.17.0 vector retrieval logic |
limit | Top 8 entries | Single batch data volume is small. Retrieving too many results increases context pressure. 8 entries cover full-process linked data for process, equipment, and exceptions for a single furnace run |
searchMode | mixedRecall | Adapts to multiple document types. Keyword matching quickly locates structured fields in process cards and furnace loading diagrams. Vector matching handles unstructured content in furnace temperature curves and metallographic reports |
usingReRank | Enabled | Heat treatment exception analysis requires precise matching of details in exception handling records. Reranking raises the sorting priority of highly relevant results, preventing irrelevant results from appearing first |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands relevant search terms such as material grades and furnace temperature parameters based on user questions. This meets professional term expansion needs in heat treatment scenarios |
collectionIds | Specified by business partition | Configure knowledge base collections by workshop, furnace type, or customer partition. This avoids introducing irrelevant data from cross-partition retrieval and improves retrieval accuracy |
Three Common Misconfigurations
- Phenomenon: Insufficient number of retrieval results, unable to cover full-process data for a single furnace run. Cause: The
limitparameter is set too low, and cross-document linked retrieval logic is not configured. This results in only single-type documents being retrieved. - Phenomenon: Retrieval response timeout, slow page loading. Cause:
collectionIdsare not specified by business partition, or lightweight retrieval configuration is not enabled. This leads to an overly large retrieval scope that exceeds hardware carrying capacity. - Phenomenon: Large model responses only mention clause frameworks, without elaborating on specific parameter content. Cause: The
similaritythreshold is set too high, ordatasetSearchUsingExtensionQueryis not enabled. This fails to retrieve detailed content of specific process clauses.
How to Verify Successful Configuration
- Submit test questions that include furnace temperature parameters and material grades. Check whether retrieval results include specific content from corresponding furnace temperature curves and process cards.
- View retrieval logs to confirm that the number of retrieved results matches the
limitparameter configuration. - Simulate high-concurrency retrieval requests to verify that response times meet business requirements, with no obvious timeouts.
- Adjust the
similarityparameter threshold, observe changes in retrieval result relevance, and confirm that the configuration takes effect.
The values provided are common starting points and should be evaluated against the reader's own samples.