What Data for This Category Looks Like
Core data for this category comes from production and testing links of battery pack, energy storage cabinet, and BMS system manufacturers. Daily documents include grouping rules, BOM lists, welding and torque records, BMS flashing records, insulation and voltage withstand reports, EOL test reports, and system configuration lists. Data sources cover cell capacity grading data, MES systems, welding equipment, torque tools, BMS software versions, and EOL test benches. Data is added or updated each time a pack or energy storage cabinet is produced, or each time a BMS software version is changed. Each individual document includes specific process parameters, test values, and version identifiers.
Constraints on Vector Models and Indexes
Multi-source heterogeneous data sources create large differences in document formats. Both structured BOMs and test reports, as well as unstructured process specification texts, are present. This requires the index system to support embedding and retrieval of mixed formats. The high-frequency update feature requires indexes to support incremental updates, avoiding inefficient full scans of hundreds of thousands of data entries each time. Professional parameters in documents, such as torque values, cell capacity grading levels, and insulation and voltage withstand values, require vector models to accurately capture semantic associations. The high update frequency also means embedding models must adapt to rapid generation requirements. In addition, the accuracy of process Q&A directly affects production safety. Retrieval result precision must strictly match business scenario standards, avoiding recall of irrelevant original device logs or incorrect configuration information.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of documents such as grouping rules and welding records, avoiding splitting process parameters with professional units into multiple chunks |
| Overlap Length | 100–200 characters | Retains cross-chunk process association information, such as the link between welding steps and subsequent torque verification |
| Similarity Threshold | 0.75–0.85 | Filters low-match non-process documents, avoiding recall of irrelevant original device logs or redundant test data |
| Single Search Limit | Top 10 entries | Controls the volume of returned information, adapts to the short context requirements of process Q&A, and avoids exceeding model processing capabilities |
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. |
defaultToken | Calibrate using representative process documents | This is the default token count used for direct text splitting; inspect sample chunks. It does not set vector dimensions or LLM context. See the v4.17.2 model-field reference |
Three Common Configuration Mistakes
- Search results include a large number of irrelevant original device logs, and target process specifications are not retrieved. The cause is failure to set a reasonable similarity threshold, allowing low-match non-standard documents to be included in the recall range.
- Single search takes more than 10 seconds after knowledge base indexing is complete, leading to low retrieval efficiency. The cause is failure to enable incremental index mode, with each query performing a full scan of hundreds of thousands of production data entries and causing slow response times.
- Professional parameter associations are lost in embedded text, such as the front and rear level information of cell grouping rules being split. The cause is overly short chunk length, splitting cross-paragraph process logic into independent chunks and destroying semantic integrity.
How to Confirm Proper Configuration
- Randomly select 10 process keywords, such as "cell grouping rules" and "welding torque requirements", run a retrieval, and check whether returned document types belong to target process documents.
- View index task running logs, confirm that incremental updates only process newly added cell capacity grading data or BMS version change files, and do not reindex old data.
- Test recall results before and after adjusting the similarity threshold, compare matching degree differences, and confirm that the threshold setting meets the precision requirements of the business scenario.
- Verify that
maxTokenis within the selected embedding model/service input limit and thatdefaultTokenproduces expected chunks. Budget retrieved text and conversation context separately against the selected LLM's documented limit.
The values provided are common starting points and should be measured against the reader's own samples.