What Data for This Category Entails
This category covers the full production workflow for air conditioners, refrigerators, washing machines, and small home appliances. Data sources include ERP/MES systems, component batch ledgers, tightening and refrigerant charging equipment, safety compliance testers, performance lab data, and after-sales fault records. Daily documents include configuration BOMs, assembly work instructions, refrigerant charging records, electrical safety test reports, random performance inspection reports, and rework records. Fields include component batch numbers, tightening torque values, charging volumes, test voltages, and rework reasons. Updates follow per-unit product, per-shift team, or random inspection schedule cadences.
Constraints Imposed on Knowledge Base Retrieval and Recall
Multi-model mixed-line assembly scenarios require retrieval results to strictly match the product model at the current workstation, to avoid interference from cross-model documents. Frequently updated per-unit or per-team data requires the retrieval pipeline to quickly locate associated documents by batch or workstation dimensions. Multiple document types including safety compliance, performance, and rework records require differentiated recall priority, to ensure core process and safety documents are returned first. Consumer-facing requirements demand retrieval results cover functional, appearance, and compliance details, to avoid omissions from single-dimensional searches.
Configuration Settings
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Balances recall accuracy and coverage, and supports precise matching for mixed-line multi-model scenarios |
limit | Top 8–12 results | Supports recall needs for multiple document types at a single workstation, and prevents context overflow from excessive returned entries |
searchMode | mixedRecall | Combines semantic matching and precise field matching, suitable for documents with structured fields such as BOMs and test reports |
usingReRank | Enabled, re-rank Top 3 results | Prioritizes documents most relevant to inspection and audit scenarios, and reduces interference from mixed-line product models |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms related to inspection and audit, and covers wording differences such as "tightening torque" and "bolt moment" |
collectionIds | Configured by product model group | Limits retrieval scope to the knowledge base set corresponding to the current workstation’s product model, and prevents cross-model documents from being included |
Three Common Misconfigurations
- When calling the chat API, an error "No knowledge base selected" is returned, but retrieval works in the debug preview interface. The cause is that request parameters do not correctly pass
collectionIds, or the deployed knowledge base is not bound to the permission scope of the current application. - When importing content from a web link that requires login, document body content cannot be obtained. The cause is that no crawler access credentials are configured for the knowledge base, so FastGPT cannot bypass login restrictions to access target content.
- Returned documents do not match the product model of the current workstation. The cause is that
collectionIdsare not configured by product model group, or thesimilaritythreshold is set too low, leading to recall of documents from unrelated product models.
How to Verify Proper Configuration
- Manually pass the specified
collectionIdsto initiate retrieval, and confirm that returned results only include documents for the corresponding product model. - Adjust the
similaritythreshold, observe changes in the matching precision of recall results, and confirm the threshold falls within a reasonable range. - After enabling
usingReRank, confirm that the returned results are sorted to prioritize core needs of the current inspection and audit scenario. - Pass the
datasetSearchUsingExtensionQueryparameter when calling the API, and confirm that retrieval results cover synonyms and alternative wordings.
The values listed above are common starting points, and should be evaluated against the user’s own samples.