What this category's data looks like
This category covers the full production process of hot rolling, cold rolling, section steel, and steel pipe. Daily documents include rolling plans, heating schedules, rolling specifications, plate shape and thickness records, surface quality reports, and quality certificates. Data comes from slab/billet batches, heating furnaces, mill automatic control equipment, online thickness and width measuring devices, surface inspection systems, and MES systems. Updates happen in real time per coil or per steel billet. Document fields include steel batch number, heating temperature, rolling reduction, plate shape deviation, and more. Units include degrees Celsius, millimeters, meters per second, and others.
What constraints do these characteristics impose on knowledge base retrieval and recall?
The real-time coil-level updates, multi-type documents, and batch association features of this category create multiple constraints for knowledge base retrieval and recall. First, each steel coil has independent production data. Retrieval must support limiting scope by steel batch number to prevent cross-batch data mixing. Second, documents include structured tables and non-specification text. Retrieval logic must adapt to multiple formats. The real-time update feature requires retrieval to cover newly stored production data, while handling index update pressure from high-frequency incremental synchronization. Additionally, production error risks require recall results to accurately match current batch process parameters, without missing key details.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category has high precision requirements for process parameters. Filter out irrelevant documents with insufficient matching scores to avoid interfering with process Q&A |
limit | Top 8–12 entries | Production documents per batch are concentrated in number. Control the number of recalled entries to fit the model context window and avoid redundant information |
usingReRank | Enabled | This category of documents contains a large number of numerical parameters. Reranking optimizes the sorting accuracy of matching results and improves the accuracy of process Q&A |
datasetSearchUsingExtensionQuery | Enabled | Supplement limiting terms such as batch number and steel grade by extending user questions to narrow the retrieval scope to the target production batch |
collectionIds | Configured by production workshop/steel grade | This category of documents is divided by production line. Specifying the retrieval scope via collectionIds avoids mixing cross-workshop data |
Three Common Misconfigurations
- Issue: Retrieval results are not limited to the target steel batch’s documents, and process data from other batches is included. Cause: No target production line knowledge base collection is specified via
collectionIds, and no batch number limiting condition is added via extended queries. - Issue: After importing CSV-formatted plate shape and thickness records, the AI only recalls some table entries and fails to cover all content. Cause:
datasetSearchUsingExtensionQueryis not enabled, and no retrieval mode adapted to structured tables is configured, leading to incorrect matching of some fields. - Symptom: Plugin confirmation fails with a MongoServerError. This error alone does not establish that MongoDB 4.4.29 is incompatible with FastGPT; inspect the complete logs, the database request issued by the plugin, and account permissions. The v4.17.2 deployment config defaults to MongoDB 5.0.32 and documents 4.4.29 as a fallback for CPUs without AVX (see the official deployment configuration).
How to Verify Proper Configuration
- Enter a test query that includes the specified steel batch number, and verify that recall results only include documents associated with that batch, such as rolling plans and heating schedules.
- Upload a structured table document, query any parameter item in the table, and confirm that all relevant table entries are included in the recall results.
- View the knowledge base configuration, confirm that
collectionIdsis bound to the current production line’s document collection to avoid cross-production line data mixing. - Test retrieval results with
usingReRankenabled and disabled, and confirm that reranking optimizes the sorting logic for process parameter matching.
The values provided are common starting points and should be evaluated against the reader’s own samples.