What the data for this category looks like
Business data for the general mechanical equipment assembly category mainly comes from PLM, ERP, parts inspection data, assembly tool collection, test benches, MES, 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 updates are triggered per individual piece of equipment or each order. Document structures typically include fields such as material codes, torque values, clearance parameters, and inspection results. Units include millimeters, newton-meters, hours, and others.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Data for this category updates per individual piece of equipment or order. The knowledge base must support associating data by order or equipment dimension to avoid mixing data across orders. There are diverse document types with large field differences. Retrieval must adapt to the field formats of different documents. For example, the material code field in BOMs and the numerical field in torque records require separate matching query logic. Data update frequency is high. Recall results must link to the latest assembly and inspection records in real time. Retrieval scope must also be controlled to avoid interference from redundant data. Additionally, the multi-variety order-based assembly characteristic requires retrieval results to distinguish exclusive configurations for different orders, ensuring recalled content matches the current assembly task.
The following configurations are based on FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the precision requirements for numerical fields such as torque and clearance, and avoids recalling irrelevant assembly documents |
limit | Top 6 entries | Matches the number of multiple types of documents for a single piece of equipment, and prevents the context from being too long to exceed the model's processing limit |
searchMode | embedding | Adapts to mixed retrieval of multiple types of documents, covering content in different formats such as process cards, BOMs, and inspection records |
usingReRank | Enabled | Performs secondary sorting on recall results to filter out low-relevance recall items caused by field differences |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms, and supplements retrieval dimensions for exclusive fields such as order numbers and equipment models |
collectionIds | Document collection ID associated with the order | Limits the retrieval scope to only the assembly data of the current order, avoiding interference from cross-order data |
Three common configuration mistakes
- Phenomenon: Zero results are returned after retrieval, or the interface prompts "no matching content". Cause: The current order's document collection ID is not filled in
collectionIds, causing the retrieval scope to exceed the assembly data of the target order. - Phenomenon: Recall results include torque records or clearance adjustment data from other orders. Cause:
datasetSearchUsingExtensionQueryis not enabled, and exclusive identifiers such as order numbers and equipment models are not added to the query terms, causing cross-order data to be recalled. - Phenomenon: The relevance of retrieval results fluctuates greatly, and some low-relevance documents are recalled first. Cause:
usingReRankis not enabled, only relying on the results of initial semantic retrieval, and no secondary sorting and filtering are performed on the recall results.
How to confirm the configuration is complete
- Check the collection binding configuration of the knowledge base, confirm that
collectionIdsis associated with the assembly document collection corresponding to the current order. - Initiate a test retrieval, check whether the number of returned results meets the
limitsetting requirements. - Verify the relevance of the retrieval results, confirm that the similarity falls within the preset
similarityrange. - Check whether the reranking function is enabled, confirm that the sorting logic of the recall results meets the business requirements.
The values provided are common starting points and should be measured against samples specific to the relevant deployment.