What This Category of Data Looks Like
The data for this category comes primarily from raw material quality certificates, heating furnace systems, press collection data, MES systems, non-destructive testing instruments, and customer forging drawings. Daily stored documents include forging process specifications, cutting lists, heating furnace logs, forging process cards, mold life records, heat treatment work orders, and ultrasonic testing reports.
Data updates follow the rhythm of individual pieces, batches, or furnace runs. Different documents correspond to attributes of distinct production links. Fields include forging part numbers, heat lot numbers, process parameters, test results, and more. Units vary by document type, with no unified fixed unit system.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The data update rhythm for this category is tightly linked to production batches. Retrieval and recall must adapt to high-frequency document updates triggered by production nodes. Multiple types of heterogeneous documents coexist, including process parameters, test reports, and production records. This requires the retrieval system to support precise matching across document types.
Data is tied to unique identifiers such as forging part numbers and heat lot numbers. This ensures retrieval results correspond strongly to the current production batch. Some testing documents contain mixed structured and unstructured content. This requires proper control of single-document segmentation granularity to avoid retrieval redundancy, while meeting real-time generation needs for maintenance Q&A.
Configuration Settings
This configuration applies to FastGPT v4.17.0
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Aligns with semantic matching accuracy for forging-specific documents, filters low-match irrelevant records, and balances retrieval accuracy and coverage |
limit | Top 8 entries | Matches the number of associated documents for a single batch of forging production, balances recall coverage and context length, and avoids exceeding model input limits |
searchMode | mixedRecall | Supports hybrid retrieval of structured parameters (such as heat lot numbers, mold numbers) and unstructured text (such as process specifications, maintenance records), adapting to multiple types of heterogeneous documents |
usingReRank | Enabled | Re-ranks recall results, optimizes sorting accuracy for specialized documents such as mold life records and ultrasonic testing reports, and improves the reference value of maintenance Q&A |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands professional terms related to inspection and maintenance, linking "mold wear" to relevant documents such as "mold life records" and "forging process cards" |
collectionIds | Divided by production workshop/forging part type | Limits retrieval scope to forging documents from the corresponding workshop, avoids interference from cross-category data, and aligns with the business structure of forging plants with multiple workshops |
Three Common Mistakes
- Phenomenon: Retrieval results include tabular documents, but the final generated maintenance Q&A output does not display table formatting. Cause: The structured content retention configuration for knowledge base retrieval is not enabled, or table-formatted text is incorrectly filtered during the re-ranking stage.
- Phenomenon: After uploading a large-volume ultrasonic testing report, retrieval response takes too long to return results in a timely manner. Cause: Reasonable chunking parameters for large documents are not adjusted, or retrieval timeout thresholds that meet business needs are not set.
- Phenomenon: Retrieval results mix documents from other workshops or different forging part types, and cannot accurately match the current production batch requirements. Cause: The
collectionIdsparameter is not configured correctly, and the scope of the knowledge base collection for retrieval is not limited.
How to Confirm the Configuration Is Correct
- Upload a mixed document containing forging process specifications and ultrasonic testing reports, perform a retrieval related to equipment inspection, and check that the recall results include both types of documents and the sorting aligns with business needs.
- Enter a query statement containing a heat lot number or forging part number, check that the retrieval results only return documents from the corresponding batch, and verify that the
collectionIdsconfiguration is effective. - Trigger retrieval for a large-volume document, confirm that the response time meets the real-time requirements of the production link, and verify the rationality of the
limitandsimilarityconfigurations. - Check whether table formatting is retained in the retrieval results, and verify that the structured content display configuration is correct.
The values provided are common starting points and should be measured against the reader's own samples.