Data Profile for This Category
Business data for general mechanical equipment assembly scenarios comes from PLM/ERP systems, parts inspection databases, assembly tool collection terminals, test benches, MES systems, and customer technical agreements. Daily generated documents include assembly process cards, multi-level BOMs, torque records, clearance adjustment records, commissioning reports, factory inspection sheets, and packing lists.
Data updates run per single device or single order. Each document includes clear fields such as model, serial number, torque value, clearance parameters, and inspection results. Units include newton meters, millimeters, megapascals, and other common engineering units.
Constraints on Knowledge Base Retrieval and Recall
The nested structure of multi-level BOMs requires precise matching of part numbers and assembly nodes during retrieval. This avoids invalid results from fuzzy matching.
The high-frequency update rhythm for single devices or orders requires knowledge base incremental sync speed to match production cycles. This prevents data lag from affecting audit efficiency.
Numeric data with clear engineering units such as torque and clearance needs consistent unit retention during retrieval. This avoids matching failures caused by unit conversion deviations.
Mixed structured and unstructured documents from multiple sources need unified chunking rules to adapt to subsequent recall logic. This avoids format differences interfering with retrieval accuracy.
Supplier audits associate part inspection records with full assembly process data. Recall results cross-reference complete information for corresponding batches across documents, supporting full-process traceability.
Configuration Settings
The following configurations are based on FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Balances recall accuracy and coverage. Avoids too low values introducing irrelevant assembly documents, and too high values missing key clearance and torque parameters. |
limit | Top 8–12 entries | Documents related to single device assembly usually do not exceed 10 copies. Excessive results exceed context window limits and affect generation quality. |
searchMode | mixedRecall | Adapts to mixed retrieval needs for structured BOM fields and unstructured process text in this scenario. |
usingReRank | Enabled | Multi-level BOMs and assembly records have high semantic similarity. Re-ranking improves the sorting priority of core documents. |
datasetSearchUsingExtensionQuery | Enabled | Supplements assembly domain synonyms such as "pre-tightening torque" and "torque pre-tightening" to improve term matching recall rate. |
collectionIds | Exclusive collection IDs divided by business domain | Isolates assembly documents and supplier inspection documents, avoiding cross-domain retrieval interfering with audit processes. |
Three Common Misconfigurations
- Deleting a collection with a large number of assembly documents triggers a
timeout of 60000ms exceedederror. Batch deletion policies are not configured, so the number of documents processed in a single operation exceeds the system's default carrying limit. - Some images fail to display during knowledge base question answering. Chunking operations do not extract image storage paths or alt text, so retrieval results cannot associate with image resources.
- Retrieving BOM documents only recalls text content, and fails to match part number fields. Special retrieval configuration for structured fields is not enabled, so only plain text content is used for matching.
How to Verify Configuration Correctness
- Enter the model of a single device. Check if recall results include the corresponding batch's assembly process cards, BOMs, and inspection records. Adjust the
similarityvalue to match required recall accuracy. - Run a batch delete collection operation. Monitor logs for timeout errors. Adjust the document count threshold for batch processing to match system performance.
- Test retrieval of documents with images. Confirm the question answering interface loads associated images normally. Verify chunking configuration correctly extracts image metadata.
- Enter assembly domain terms such as "coaxiality adjustment". Check if recall results include documents related to extended synonyms. Verify the
datasetSearchUsingExtensionQueryconfiguration works.
The values provided here are common starting points, and should be evaluated against the reader's own samples.