What the data for this category looks like
This category's data primarily comes from order design systems, sheet material and wood batch ledgers, CNC cutting and drilling equipment logs, ERP/MES systems, and moisture content testing equipment. Daily documents include order disassembly drawings, cutting lists, edge banding and drilling programs, coating process documents, hardware material lists, and assembly and packaging records. Data updates are triggered per order, daily, or by material batch. Fields include panel size, batch number, process parameters, test results, and similar items. There are no uniform fixed units, as they vary based on document type.
What constraints these characteristics impose on the vector models and indexes workflow
Custom order-driven multi-panel attributes require indexes to support associative retrieval by batch and order dimensions. The large number of panels and high volume of data per single order require controlling storage and retrieval efficiency for single-batch indexes. Documents contain mixed text and image content (order disassembly drawings, process documents), requiring adaptation to multimodal vector processing. Frequently updated batch data requires support for incremental index construction to avoid excessive time spent on full reconstruction. Hole position and hardware information associated with on-site installation require vector retrieval to accurately match field association relationships.
How to set configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the mixed text and image content length of order disassembly drawings and cutting lists, avoiding breakage of critical information within chunks |
| Overlap Length | 100–150 characters | Retains associated information between adjacent chunks, preventing fields such as panel size and batch number from being split |
| Similarity Threshold | 0.75–0.85 | Filters low-match retrieval results, adapting to the precise matching requirements of multi-panel items in a single order |
| Single Search Limit | Top 20 entries | Controls the volume of retrieval returns, avoiding excessive redundant data interfering with defect analysis |
defaultToken | 4096 | Adapts to the vector generation length of long texts such as coating processes and assembly records |
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. |
Three common mistakes
- A prompt indicates vector model configuration failed, with status code 400 returned. The cause is failure to adapt to the mixed text and image documents of this category, selecting a vector model that only supports plain text, and not enabling multimodal vector support.
- The number of retrieval results is far lower than expected, with only 1-2 matching entries returned. The cause is that the similarity threshold is set too high, filtering some qualifying panel batch association data.
- Index construction takes too long, exceeding the preset waiting time. The cause is not enabling incremental index mode, performing full reconstruction on all order data, and not splitting index tasks by material batch or order dimension.
How to confirm proper configuration
- Upload a single order's order disassembly drawing and cutting list, check that the vector generation status shows completed with no error prompts.
- Enter the search term "panel size deviation", verify that the match degree of returned results falls within the preset similarity threshold range.
- Batch import test records for 10 sheets of the same batch, check that the index completes incremental updates by batch with no full reconstruction prompts.
- View the vector model call logs, confirm that the values of
defaultTokenandmaxTokenmeet the configuration requirements, with no text truncation prompts.
The values provided are common starting points and should be measured against the reader's own samples.