What this category of data consists of
The data sources for this category include raw material quality certificates, heating furnace systems, press collection data, MES, non-destructive testing instruments, and customer forging drawings. Daily documents cover forging process specifications, cutting cards, heating furnace records, forging process cards, mold life records, heat treatment orders, and ultrasonic testing reports. Data updates follow per piece, per batch, or per furnace heat cycles. A single data entry includes fields such as material furnace batch number, forging batch number, process parameters, and testing results, using professional units including temperature, pressure, and forging ratio.
Constraints imposed by this category's characteristics on knowledge base retrieval and recall
This category of data has a high update frequency and is split by batch and furnace heat. Retrieval must support precise filtering of results by furnace batch number and forging batch number. Daily documents include multiple content types such as process specifications and testing reports. A mixed recall logic for structured parameters and unstructured text must be supported. Data sources cover the entire production process. Retrieval results must associate information from raw materials, production processes, and testing links to support complete drafting of process documents and abnormal reports. The accuracy of professional parameters directly affects subsequent drafting quality. The field matching degree of recall results must meet business specifications.
Note: This configuration applies to FastGPT v4.17.0.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Filters low-correlation process parameters and testing results to avoid mixing irrelevant content when drafting abnormal reports |
limit | Top 10 entries | Balances the coverage of full-link information required for process and abnormal reports and retrieval efficiency |
searchMode | mixedRecall | Adapts to the content characteristics of this category where structured process parameters and unstructured documents coexist |
usingReRank | Enabled | Re-ranks recall results by relevance to the current drafting task to optimize result display order |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms to cover synonymous expressions of professional terms such as forging ratio and fiber direction |
collectionIds | Bind the knowledge base collection corresponding to the current forging workshop | Limits the retrieval scope to the exclusive data of the current production link to avoid interference from cross-workshop data |
Three common misconfigurations
- Symptom: After importing a knowledge base via mongorestore, retrieval returns no matching results and an empty list. Cause: The target collection ID is not bound in the retrieval configuration, or the import operation did not trigger index rebuilding, resulting in data not being indexed by the system.
- Symptom: Recall results contain a large amount of homogeneous content, and older documents are displayed first. Cause: The
usingReRankre-ranking function is not enabled, and the associated logic for sorting by document update time is not configured, resulting in content closer to the current drafting date not being prioritized. - Symptom: After performing a batch retraining operation, retrieval results do not change as expected. Cause: The extended retrieval term configuration is not updated synchronously, or the training does not cover all forging process and testing documents, resulting in some data not being re-embedded.
How to verify the configuration is correct
- Manually input a known existing furnace batch number, and verify that retrieval results include the process specifications and testing reports corresponding to that furnace number.
- Access the retrieval configuration panel, and confirm that the switch states of
usingReRankanddatasetSearchUsingExtensionQuerymatch the expected settings. - Import a newly added forging testing report, run a retrieval, and confirm that the new content can be recalled normally.
- Compare retrieval results from different batches, and confirm that the batch filtering logic functions correctly.
The values provided are common starting points and should be evaluated against local sample data.