What Data for This Category Looks Like
This category covers daily production and after-sales documents for air conditioner, refrigerator, washing machine and small appliance manufacturers. Documents include bill of materials (BOM), assembly work instructions, refrigerant charging records, electrical safety test reports, performance spot check reports and repair records. Data sources cover ERP/MES systems, component batch data, tightening and charging equipment, safety compliance testers, performance laboratories and after-sales fault data. Update cycles run per single device, per work shift team, or according to spot check plans. Document structures include fixed fields such as component model, process parameters, test results and repair reasons. Some documents include original values collected by equipment and batch identifiers.
Constraints for Vector Models and Indexes
Mixed-line assembly production creates a business scenario where multiple model documents coexist. Indexes must have precise model differentiation capabilities to avoid incorrect matching of cross-model process documents. High-frequency single-device or team-level data updates require indexes to support incremental update logic. This reduces resource consumption from full index rebuilding. Strongly compliant documents such as safety compliance and performance spot check reports include structured numerical values and unstructured text. Vector models must adapt to mixed input formats. A need exists to associate after-sales repair records with front-end process documents. Indexes must support cross-document associated retrieval, while controlling single-batch index scale to ensure retrieval speed.
How to Set Configurations
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to semantic integrity of single-page documents such as assembly work instructions and test reports, avoids semantic breaks across paragraphs |
| Overlap Length | 50–80 characters | Retains semantic connection between adjacent chunks, adapts to continuous parameter description paragraphs in process documents |
| Similarity Threshold | 0.72–0.85 | Calibrated based on actual testing. Ensures document differentiation in mixed-model production scenarios, avoids incorrect matching of similar process parameters across different models |
| Single Search Limit | Top 10 results | Controls the scale of retrieval results, adapts to the need for quick positioning when searching process documents |
maxToken | 4096 | Adapts to vector generation for long documents such as safety compliance test reports, avoids semantic loss caused by content truncation |
defaultToken | 1024 | Adapts to vector generation for short documents such as repair records, balances generation efficiency and semantic integrity |
Three Common Misconfigurations
- Errors occur after configuring the vector model, and valid indexes cannot be generated. The cause is failure to adjust input formats for structured process parameters, leading to the model being unable to correctly parse numerical fields.
- The number of retrieval results does not match expectations, with only a small number or no matching results returned. The cause is that the similarity threshold is set too high, filtering out valid results that match process parameters.
- The PDF enhanced parsing function fails to trigger properly for locally deployed versions. The cause is failure to confirm the local non-commercial version permissions for FastGPT v4.17.0, and incorrect configuration of parsing dependencies.
How to Confirm Proper Configuration
- A single assembly work instruction is uploaded, and chunked text fragments are verified to cover complete process steps, with no obvious semantic truncation.
- A search term including a specific component model is input, and the document type and model of retrieval results are checked to confirm matching, with the filtering effect of the similarity threshold meeting expectations.
- Multiple process documents of different models are uploaded, and retrieval results are verified to only return relevant content for the corresponding models, with no cross-model incorrect matching.
- Vector generation logs are reviewed, and the
maxTokenanddefaultTokenconfigurations are confirmed to be correctly applied, with no parameter conflict errors.
The values provided are common starting points, and should be evaluated against samples relevant to the specific deployment.