What Data for This Category Looks Like
Daily documents for this category include configured BOMs, assembly work instructions, refrigerant charging records, electrical safety test reports, performance sampling inspection reports, and rework records. Data sources include ERP/MES systems, parts batch ledgers, tightening and charging equipment logs, safety tester data, performance laboratory test results, and after-sales fault feedback databases. Data update frequencies vary: records for individual products are generated in real time, shift-level documents are updated per shift, and sampling reports are generated according to established plans. Document fields include part numbers, process parameters, test values, batch identifiers, timestamps, and additional relevant fields. Some documents include physical units: refrigerant charging volume is measured in grams, and tightening torque is measured in newton meters.
Constraints on Vector Models and Indexes
This category’s data characteristics create multiple constraints for the vector model and indexing workflow. Multi-source, heterogeneous data sources include equipment logs, system ledgers, and test reports. These require unified embedding rules to avoid semantic deviation in identical fields across different sources. High-frequency real-time individual product records are stored alongside low-frequency sampling reports. A layered indexing strategy is needed to differentiate priority levels for real-time queries and batch recall. Documents carry precise process parameters and physical units. Vector modeling must retain semantic details of all fields to prevent retrieval errors caused by confusion between units and parameters. In mixed-line assembly scenarios with multiple product models, indexes must support quick filtering by product model. This prevents incorrect matching of cross-model process documents.
How to Configure
The following configurations are based on FastGPT v4.17.0. The values for defaultToken and maxToken must be adapted to the long document characteristics of this category.
| Configuration Item | Recommended Range/Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | This category’s documents include long sections of process steps and test data. This length preserves complete semantic units and avoids cutting process logic. |
| Overlap Length | 100–150 characters | Process specifications have linked consecutive steps. Overlapping chunks preserve context coherence and prevent loss of step connection information during retrieval. |
| Similarity Threshold | 0.75–0.85 | This range distinguishes process parameter differences between the same model and different batches. It avoids recalling irrelevant documents with low matching scores. |
defaultToken | 8192 | Adapts to the embedding needs of this category’s long documents. It ensures complete process parameters and test results can be correctly vectorized. |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
| Single Search Limit | Top 8 results | Mixed-line assembly scenarios require quick location of process documents for the corresponding model. Too many results increase on-site judgment costs. |
Three Common Misconfigurations
- Symptom: The knowledge base remains in the "Indexing" status for an extended period after creating a new data index, with no progress updates. Cause: The configuration does not account for this category’s high-frequency real-time individual product records. The index refresh trigger frequency is set too high, which consumes excessive system resources.
- Symptom: Retrieval returns no matching content after applying custom index configurations, or matching results are unrelated to process-related questions. Cause: No semantic mapping is configured for this category’s process parameters with physical units. This leads to loss of association between units and parameters during vector embedding.
- Symptom: Existing data cannot be loaded after connecting a custom vector database. Cause: Fields including batch numbers, tightening torque, and refrigerant charging volume for this category have not been converted to the vector storage format required by FastGPT. This prevents the database from recognizing field semantics.
How to Verify Proper Configuration
- Upload the assembly work instruction and test report for a single home appliance product. Run a process-related retrieval. Check if returned results include process content for the corresponding product model.
- Review FastGPT indexing task logs. Confirm that all multi-source data for this category has completed vectorization and index creation.
- Enter a process query that includes physical units, such as "Refrigerator electrical safety test standards". Verify that retrieval results match information in the corresponding documents.
- Test retrieval result coherence across different chunk configuration settings. Confirm that process steps are not overly fragmented.
The values provided here are common starting points. They should be evaluated against the reader’s own samples.