What this category of data looks like
This category of data comes from material grade standards, customer technical requirements, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers and metallographic testing equipment, and manufacturing execution systems. Data is updated per heat and per batch. Daily documents include heat treatment process cards, furnace loading layout diagrams, real-time furnace temperature curves, hardness and metallographic test reports, equipment inspection forms, and exception handling sheets. Data fields include furnace temperature range (unit: ℃), furnace loading weight (unit: kg), hardness value (unit: HV or HRC), holding time (unit: minute), etc. Most documents are structured tables and time-series curve files.
What constraints do these characteristics impose on knowledge base retrieval and recall
The data characteristics of this category impose multiple constraints on knowledge base retrieval and recall. Documents generated by multi-source equipment and production systems include structured tables, time-series curves, and plain text. This requires the retrieval pipeline to support vector modeling and associated recall for multiple formats. Data is updated frequently per heat and per batch. This requires the recall process to support low-latency incremental synchronization, to avoid retrieval results that do not match current production requirements due to data lag. Fields use industry-specific units. Retrieval must match process conditions associated with units, to avoid invalid cross-unit matches. There are strong associations between different documents. For example, furnace loading layout corresponds to furnace temperature curves, and process cards correspond to hardness test reports. Recall must support cross-document context association, to ensure the completeness of retrieval results.
The following configurations are based on FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.6–0.7 | Matches the precise parameter requirements of heat treatment processes, filters low-relevance general process documents |
limit | Top 6–10 entries | The number of ledger documents associated with a single heat is limited, controlling the number of recalled entries to avoid context redundancy |
searchMode | fullTextRecall | Adapts to associated retrieval of multi-format documents, covering cross-type data such as furnace loading diagrams and furnace temperature curves |
usingReRank | Enabled | Reorders initial recall results by process relevance, correcting sorting bias from initial retrieval |
datasetSearchUsingExtensionQuery | Enabled | Expands process-related keywords based on user questions, such as expanding "quenching requirements" to cooling speed and holding time |
collectionIds | Specified by production line/workshop | Restricts retrieval scope to the heat treatment knowledge base of the current production line, excluding historical data from unrelated production lines |
Three common mistakes
- Phenomenon: A large number of irrelevant general heat treatment process documents are returned during retrieval, and ledger data for the target heat cannot be located. Cause: The
similaritythreshold is set too low, failing to match the precise parameter requirements of heat treatment processes, resulting in recall of low-relevance content. - Phenomenon: Duplicate entries appear in retrieval results after the same furnace temperature curve file is uploaded repeatedly to the knowledge base. Cause: Document deduplication configuration is not enabled, or the deduplication rule does not match the unique identifier specific to heat treatment documents (such as heat number, process number).
- Phenomenon: When retrieving furnace loading layout requirements, only furnace loading diagram documents are recalled, and corresponding furnace temperature curves and hardness reports are not included. Cause: Extended retrieval or re-ranking functions are not enabled, making it impossible to establish process associations across documents.
How to confirm the configurations are set correctly
- Upload a single heat treatment process card and furnace temperature curve, initiate retrieval for that heat, and check whether the recall results include the target documents.
- Adjust the
similaritythreshold, compare recall results under different thresholds, and confirm that they meet the parameter matching accuracy requirements of the current production line. - Enable re-ranking and extended retrieval functions, initiate a retrieval request for cross-document association, and confirm that the recall results include multiple types of associated documents.
- View the knowledge base synchronization logs, confirm that frequently updated heat data has completed incremental synchronization with no lag.
The values provided are common starting points and should be measured against the reader's own samples.