What the Data for This Category Looks Like
Business data for general mechanical equipment assembly scenarios typically originates from PLM/ERP systems, parts inspection records, assembly tool collection data, test benches, MES systems, and customer technical agreements. Daily documents include assembly process cards, BOM lists, torque records, clearance adjustment records, commissioning reports, factory inspection sheets, and packing lists. Data is updated once per single device or order. Document fields include component numbers, torque values, clearance parameters, inspection dates, qualification status, and other fields. Some documents include quantifiable fields with clear units such as physical dimensions and load test values.
Constraints on Knowledge Base Retrieval and Recall Imposed by These Data Characteristics
Three core constraints are imposed on retrieval and recall by the data characteristics of this category. First, the update rhythm tied to single devices or orders requires the retrieval system to support precise association and recall of corresponding documents by order number or device model, to avoid interference from invalid data across orders. Second, documents contain quantifiable fields such as torque, clearance, and load test values, so the system needs to support matching retrieval for numeric parameters, while also supporting associated matching between text keywords and quantifiable parameters. Third, the business model of multi-variant custom assembly requires the knowledge base to support mixed retrieval of general process cards and customized order documents, while adapting to heterogeneous data formats from multiple sources such as PLM/ERP and MES.
Configuration Settings
The following configurations are based on the parameter logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to mixed matching of quantifiable parameters and process text for this category, avoids recalling irrelevant low-relevance documents |
limit | Top 8–12 entries | Covers core information of multiple document types such as assembly records, BOM, and inspection sheets for a single device, avoids context overflow |
searchMode | mixedRecall | Balances text keyword matching (such as order numbers) and vector semantic matching (such as clearance adjustment abnormalities), adapts to multiple document types |
usingReRank | Enabled | Performs secondary ranking on recall results, filters low-quality documents caused by accidental keyword matching, improves the accuracy of defect analysis |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval queries, supplements associated retrieval for fields such as order numbers and device models, adapts to the requirement of recall by order |
collectionIds | Specified by assembly document classification | Only recalls documents from target collections such as assembly process cards, BOM, and inspection records, avoids mixing in non-assembly data |
Three Common Misconfigurations
- Phenomenon: Uploaded assembly process card images are not included in retrieval results, only plain text content is returned. Cause: The knowledge base’s image vectorization configuration is not enabled, or documents with images are not correctly uploaded to the specified assembly document collection.
- Phenomenon: Only a single result is recalled and sent to the large language model, failing to cover multiple associated BOMs and clearance adjustment records. Cause: The
limitparameter is not adjusted to a reasonable range, or multi-result merging processing logic is not configured. - Phenomenon: A large number of non-assembly parts processing documents appear in retrieval results. Cause: The
collectionIdsparameter is not specified, causing the retrieval scope to cover non-target collection data across the entire knowledge base.
How to Verify Successful Configuration
- Access the knowledge base configuration page of the FastGPT application, verify that the
similarityparameter falls within the 0.75–0.85 range, and that thelimitparameter is set to 8–12 entries. - Upload an assembly process card containing torque records, initiate a retrieval for "clearance adjustment abnormalities", and confirm that the recall results include the document’s text content and quantifiable parameters.
- Initiate a retrieval for a specific order number, and confirm that only BOM, inspection sheets, and other documents associated with that order are returned, with no non-target data from other orders included.
- After enabling
usingReRank, compare the order of retrieval results before and after activation, and confirm that the ranking of highly relevant assembly documents is improved.
The values provided here are common starting points, and should be evaluated against the reader's own samples.