What Data for This Category Looks Like
Data sources include ERP systems, MES manufacturing execution systems, component batch ledgers, tightening and refrigerant charging equipment logs, safety tester data, performance laboratory test reports, and after-sales fault records. Daily documents cover configuration BOMs, assembly work instructions, refrigerant charging volume records, electrical safety test reports, random performance inspection reports, and repair records. Data update frequencies cover single-unit equipment output, shift production summaries, and fixed random inspection plans. Fields include component batch numbers, tightening torque values, charging volumes, test voltages, repair reasons, and corresponding physical units.
Constraints Imposed by These Characteristics on Vector Models and Indexes
Mixed-line high-volume production requires indexes to accurately distinguish between multiple model documents, to avoid mixing documents for different appliance models. Frequently updated production and test data requires indexes to support incremental updates, to avoid performance losses from full reconstruction. Mixed structured and unstructured document types requires vector models to adapt to both parameterized fields and natural language descriptions. Audit requirements for safety and functional anomalies require retrieval recall accuracy to match business risk levels. The maxToken and defaultToken parameters must be set appropriately to fit documents of different lengths, while also considering reasonable allocation of hardware resources to support real-time audit scenarios.
How to Configure Settings
The following configurations are formulated based on the vector index module of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Balances semantic integrity and retrieval efficiency for household appliance documents, adapting to content of varying lengths such as assembly work instructions and test reports |
maxToken | 4096 token | Adapts to the input upper limit of mainstream open-source and commercial embedding models, preventing document chunks from exceeding the model's processing range |
| Similarity Threshold | 0.72–0.85 | Distinguishes subtle differences between batches of the same model, avoiding document matching errors during mixed-line production |
| Single Search Limit | Top 10 entries | Covers the number of associated records required for audits, avoiding excessive redundant results that interfere with judgment |
defaultToken | Calibrated via actual testing | Adapts to the default token limit of the selected embedding model, matching the average length of household appliance documents |
| Maximum Context | 3000–4000 characters | Meets the need to reference multiple adjacent process records during audits, ensuring coherent contextual semantics |
Three Common Configuration Errors
- An embedding service returns a 404 error via the interface prompt. The cause is that the API endpoint address or key permission for Baidu embedding-v1 was not configured correctly, preventing normal request initiation.
- The vector model service cannot start without a GPU environment. The cause is failing to switch to a CPU-compatible lightweight embedding model, or not enabling CPU running mode in the oneapi configuration.
- Vector index construction times out. The cause is failing to split ultra-long documents, causing
maxTokento exceed the model limit, while also failing to allocate hardware resources reasonably, and not making full use of the parallel processing capabilities of two 3090 graphics cards.
How to Confirm Proper Configuration
- Upload a typical refrigerator assembly work instruction, check whether the chunked text fully retains key field information such as model and tightening torque.
- Initiate an audit retrieval for electrical safety anomalies, confirm that the returned results include safety test records and corresponding repair data for the same batch of equipment.
- View the vector index update log, confirm that newly added production records for single units can be properly indexed, with no omissions or incorrect matches.
- Adjust the similarity threshold, verify that the trend of the number of retrieval results changing with the threshold aligns with business expectations, with neither excessive irrelevant results nor too few matching entries.
The values provided are common starting points and should be measured against the reader's own samples.