This document is compatible with FastGPT v4.17.0.
What the data for this category looks like
Daily documents for this category include assembly rules, BOM lists, welding and torque records, BMS flash records, insulation pressure resistance reports, EOL test data, and system configuration lists. Data sources include cell capacity grading data, MES system, welding equipment, torque tools, BMS software version records, and EOL test benches. Update frequency is triggered once per produced battery pack or energy storage cabinet, and updated synchronously with BMS software version changes. Document fields include cell capacity, internal resistance, material code, torque value, test voltage, software version number, and some fields have clear units.
What constraints do these characteristics impose on vector models and indexes
Data for this category is updated frequently per individual pack or cabinet, and includes both structured test data and unstructured rule documents. This requires vector indexes to support small-batch real-time updates, to avoid the time cost of full index rebuilding. Different document types have significant differences: assembly rule documents are relatively long, while welding record documents are mostly short entries. This requires adapting chunking strategies for differentiated processing. Structured fields have clear units, so the vector model must accurately identify and encode numerical text with units to avoid semantic confusion. Multi-source data is closely linked, so the index must support associating vector data from different documents using fields such as cell code and BMS version number.
How to set the configuration
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Ensures semantic integrity for long documents such as assembly rules and BOM lists, avoids truncating key clauses across chunks |
| Overlap Length | 50–80 characters | Covers associated information across chunks, prevents key content of assembly rules from being split |
maxToken | 4096 token | Matches the average length of energy storage system integration documents, ensures complete loading of semantic information from a single document chunk |
defaultToken | 1024 token | Adapts to short-entry documents such as welding and torque records, optimizes index construction efficiency |
| Similarity Threshold | Calibrated via actual testing | Must be adjusted based on the matching accuracy of insulation tests and BMS configurations, to ensure accuracy of cross-review |
| Single Recap Limit | Top 8 entries | Covers the number of associated documents for a single packaged product, avoids missing key test records |
Three common mistakes
- Symptom: The interface only supports selecting one vector model, but the business requires adaptation to multiple models, or local and server models are inconsistent. Cause: The embedding dimensions and text preprocessing rules of vector models were not unified in advance, resulting in incompatible vectors generated by different models.
- Symptom: A single file remains in the indexing state for a long time and cannot complete construction. Cause: The chunking strategy was not adjusted for short-entry documents, or a single file contains content exceeding the index limit, causing the indexing process to block.
- Symptom: Multiple uploaded data sets cannot be correctly associated with corresponding vectors. Cause: Index association rules were not configured based on document association fields such as cell code and BMS version number, resulting in a disconnect between vectors and business data.
How to verify correct configuration
- Upload a single welding and torque record document, check the index completion status, and confirm time consumption meets expectations.
- Enter a query related to assembly rules, verify that the number of returned associated documents matches the business logic.
- After replacing the vector model, verify that the cosine similarity change of vectors generated from the same document meets expectations.
- Upload multiple associated documents such as capacity grading data and assembly rules for the same cell, confirm that the index can correctly associate the two types of documents.
The values provided here are common starting points, and should be measured against the reader's own samples.