What the data for this category looks like
Daily data for the general mechanical equipment assembly category includes assembly process cards, multi-level BOMs, torque records, gap adjustment records, commissioning reports, factory inspection certificates, packing lists, and more. Data primarily comes from PLM/ERP systems, parts inspection databases, assembly tool collection terminals, test benches, MES manufacturing execution systems, and customer technical agreements. Data updates are performed per single equipment unit or single order. Corresponding documents and fields sync immediately after completing one assembly or delivering one order. Data fields include part numbers, torque values, gap parameters, test results, inspection numbers, and some fields have clear units such as newton meters, millimeters, and megapascals.
What constraints do these characteristics impose on the vector model and indexing process
This category’s data characteristics impose multiple constraints on the vector model and indexing workflow:
- The nested structure of multi-level BOMs requires retaining hierarchical relationships during chunking. Do not split cross-level material information, as this will prevent retrieval from matching complete material links.
- Data updated per single unit or order requires indexes to support incremental updates, rather than full reindexing, to avoid redundant calculation of historical data.
- Documents contain a large volume of engineering parameters with units. Vector models must adapt to these numeric fields with clear units, otherwise retrieval accuracy will decline.
- Field differences across documents are significant. For example, assembly process cards include process parameters, while packing lists include logistics information. Indexes must support multi-field mixed retrieval to cover the field types of different document types.
Configuration Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-section text length of BOMs and process cards, avoids splitting cross-level BOM nodes |
| Overlap Length | 100–150 characters | Retains contextual connections between chunks, prevents hierarchical information of nested BOMs from being split and broken |
| Similarity Threshold | 0.75–0.85 | Filters low-match engineering parameters, avoids retrieving irrelevant assembly records |
| Single Search Limit | Top 10 results | Controls the number of retrieval results, adapts to the multi-material verification scenario in tender documents |
defaultToken | 4096 | Matches the input length requirements of general engineering text, FastGPT v4.17.0 natively supports this parameter |
maxToken | 8192 | Supports contextual input for complete BOM hierarchies and test report content |
Three Common Misconfigurations
- Phenomenon: No optional options appear in the index model dropdown menu, and the interface displays a load failure prompt. Cause: The local model API address is not configured, or the model is not properly loaded into the FastGPT runtime environment.
- Phenomenon: A large number of irrelevant gap adjustment records appear in retrieval results, with lower-than-expected matching scores. Cause: The similarity threshold is set too low, failing to filter low-match non-target documents.
- Phenomenon: Empty field errors occur during the BOM verification step in tender documents. Cause: Chunk length is set too large, causing BOM hierarchical number fields to be lost during splitting, preventing vector retrieval from matching complete material information.
How to Confirm Proper Configuration
- Upload a single equipment BOM document and assembly process card, run a retrieval test, and check if returned results include matching part numbers and process parameters.
- Check the vector index update log, confirm that index synchronization for a single equipment document completes within 10 minutes, aligning with the per-order update rhythm.
- Call the configured vector model API, verify that the
defaultTokenandmaxTokenparameters take effect, with no truncation errors when inputting long text. - Adjust the similarity threshold, compare the number of retrieval results across different thresholds, and confirm the threshold setting meets the matching accuracy requirements of the business scenario.
The values listed here are common starting points and should be evaluated against the reader's own samples.