What Data Looks Like for This Category
Data for general mechanical equipment assembly scenarios typically comes from PLM/ERP systems, part inspection data, assembly tool collection terminals, test benches, MES systems, and customer technical agreements. Daily documents include assembly process cards, BOM tables, torque records, clearance adjustment records, commissioning reports, factory inspection certificates, and packing lists. Data updates occur per single equipment unit or individual order. Document fields include material codes, tightening torque values, coaxiality parameters, inspection results, and more. Some documents have clear unit constraints: for example, torque records use newton meters, and clearance adjustment records use millimeters.
Constraints Imposed on Knowledge Base Retrieval and Recall
This category has scattered data sources and frequent update cycles. The knowledge base retrieval link must support associated matching of multi-source data to avoid information loss caused by fragmented data sources. Document types include structured BOM tables and unstructured process records. The retrieval logic must meet both field matching and full-text semantic matching requirements. Documents for a single device are strongly bound to orders. Retrieval must first recall content matching the current order’s parameters. It is also necessary to address index delays caused by high-frequency updates to ensure recalled content is synchronized with the latest production data.
Configuration Settings
This page’s configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Meets the dual requirements of field matching and semantic association in assembly documents, avoiding recall of irrelevant BOM or process content |
limit | Top 6 entries | Controls the number of documents returned per retrieval round, adapting to the information density requirements of assembly ledger summary |
searchMode | mixedRecall | Balances structured field matching and full-text semantic retrieval, adapting to the mixed document format of BOM tables and process cards |
usingReRank | Enabled, retain top 3 reranked results | Filters redundant content from initial recall, prioritizing documents with higher matching degrees to assembly parameters |
datasetSearchUsingExtensionQuery | Disabled | Documents in this category are strongly bound to orders, no need for additional query expansion, avoiding introduction of irrelevant information |
collectionIds | Knowledge base collections bound to orders | Restricts retrieval scope, only recalling assembly documents associated with the current order |
Three Easy-to-Make Configuration Errors
- Phenomenon: A model call error prompt appears after uploading a 3MB PDF file to the knowledge base. Cause: No appropriate chunking rules are configured for structured documents such as BOM tables in this category, triggering model context length limits during large file parsing.
- Phenomenon: The knowledge base deployed via Docker remains in the indexing state continuously, with no progress updates. Cause: No retrieval knowledge base collection is restricted, and the system attempts to fully scan scattered assembly documents, leading to backlogged indexing tasks.
- Phenomenon: The ledger content returned by AI conversations is adjusted and organized, not strictly matching the parameter descriptions in the original documents. Cause:
datasetSearchUsingExtensionQueryis enabled, or the result organization switch is not disabled, causing the model to perform secondary processing on recalled content.
How to Confirm Configuration Is Correct
- Upload the assembly process card and BOM table for a single device, perform a retrieval test, and verify whether the returned results include documents associated with the current order.
- View the knowledge base index logs, confirm that only the collections corresponding to the configured
collectionIdsare scanned, with no redundant scanning tasks. - Input a query containing torque parameters, verify whether the matching degree of the recalled results meets the preset requirements.
- Trigger a ledger summary task, confirm that the returned content is not secondarily adjusted, and matches the field descriptions in the original documents.
The values provided here are common starting points and should be evaluated against the reader’s own samples.