What Data Looks Like for This Category
The typical entities in this category are air conditioner, refrigerator, washing machine and small household appliance manufacturers. Daily generated documents include configuration BOMs, assembly work instructions, refrigerant charging records, electrical safety test reports, performance sampling test reports and repair records. Data sources cover ERP/MES systems, component batch ledgers, tightening and charging equipment logs, safety compliance tester data, performance laboratory reports and after-sales fault databases. Data update frequency follows individual equipment, shift production lines or sampling plans. Each data entry includes dedicated fields such as equipment model, process parameters, test results and fault descriptions.
Constraints Imposed on Vector Models and Indexes
The mixed-line assembly of multiple product models requires indexes to support precise filtering by equipment model. This prevents incorrect recall of cross-model documents. Structural differences across multi-source heterogeneous documents require vector models to adapt embedding logic for different fields. For example, numeric fields like process parameters and text fields like fault descriptions need distinct embedding logic. Frequently updated production and after-sales data requires indexes to support incremental updates. This avoids the time cost of full index reconstruction. The direct consumer-facing attribute requires retrieval results to link to the latest after-sales fault data. This ensures repair recommendations align with actual user feedback.
Configuration Setup
The following configurations are adapted for the complete household appliance manufacturing scenario using FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the average length of single documents such as assembly work instructions and safety compliance reports, avoids breaking process steps or fault descriptions within a chunk |
| Overlap Length | 100–150 characters | Covers cross-chunk associated information, such as connected content between charging records and repair records, prevents key associations from being split |
maxToken | 8192 token | Matches the context limit of mainstream embedding models, can carry complete BOM, process and after-sales associated data for a single piece of equipment |
defaultToken | Counted per application | Enables application-level token counting, supports splitting consumption data by production line and equipment model |
| Single Search Limit | Top 10 results | Controls the scale of retrieval results, adapts to the rapid response needs of production sites |
| Reranking Model | dengcao/Qwen3-Rerank | Adapts to locally deployed embedding and reranking models, improves matching accuracy for fault descriptions |
Three Common Misconfigurations
- Symptom: Retrieval results include incorrect cross-model matches, returning equipment documents not from the current production line. Cause: No targeted similarity threshold was set for the multi-model mixed-line scenario of complete household appliances. This leads to recall of irrelevant documents with low matching scores.
- Symptom: Milvus fails to start during deployment, with the compose file including a large number of PostgreSQL-related configurations. Cause: A generic vector database deployment template was used incorrectly. No lightweight index deployment configuration adapted for the household appliance manufacturing scenario was used.
- Symptom: Token consumption cannot be counted per application. Cause: The application-level statistics switch for
defaultTokenwas not enabled. This prevents token data from being split by production line and equipment model.
How to Verify Proper Configuration
- Import assembly work instructions, safety compliance reports and after-sales repair records for a single equipment model. Run equipment inspection Q&A, and confirm returned results only include associated documents for that model.
- View the index storage usage of the vector database. Confirm that index data for a single piece of equipment does not exceed the preset threshold, meeting batch deployment requirements for production lines.
- Run locally deployed embedding and reranking models. Input typical fault descriptions such as "refrigerant leak", and confirm the sorting and matching accuracy of retrieval results meet expectations.
- Enable the token statistics function. View token consumption data for different production lines, and confirm statistical dimensions cover the needs of the application scenario.
The values provided here are common starting points, and should be evaluated against the reader's own samples.