What the Data for This Category Looks Like
This category of data mainly comes from automotive body panel, home appliance sheet metal, and hardware stamping workshops. It covers documents such as stamping work instructions, material utilization tables, mold inspection checklists, first/last piece records, dimensional inspection reports, and mold change records. Data sources include steel coil labels and quality certificates, punch and feeder operation data, mold management systems, MES systems, and coordinate measuring machine and inspection tool detection data. Updates are triggered by coil batches, production batches, and mold change operations. Each single update includes all documents and device collection fields associated with a single production batch.
Constraints Imposed on Vector Models and Indexes
Data update frequency triggers dynamically with production batches, so indexes must support incremental updates to avoid excessive compute resource usage from full rebuilds. Multi-source heterogeneous data includes text, numerical, and device log formats, requiring unified vector embedding adaptation logic. defaultToken and maxToken must be properly configured to meet embedding requirements for multiple field types, ensuring semantic associations across different field types are correctly captured. Single batch data volume fluctuates with mold changes, so index sharding must support dynamic scaling to prevent retrieval latency during peak periods. Additionally, strong associations exist between documents, for example mold change records correspond to mold inspection checklists. Cross-review requires retaining inter-document association indexes; independent chunk storage will break semantic associations.
Configuration Settings
This configuration is based on the feature logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the length of single-operation descriptions in stamping work instructions, avoiding semantic breaks across processes |
| Overlap Length | 150–200 characters | Covers associated fields between mold change records and subsequent first/last piece records, preventing cross-chunk semantic loss |
| Similarity Threshold | 0.75–0.85 | Filters duplicate mold inspection checklist entries while retaining reasonable differences in material utilization tables across batches |
| Single Search Limit | Top 10 entries | Adapts to the number of associated documents for a single production batch in stamping workshops, avoiding redundant retrieval results |
defaultToken | 2048 | Fits the embedding length requirements for numerical and descriptive fields in dimensional inspection reports |
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 Misconfigurations
- Phenomenon: Duplicate document chunks appear after knowledge base index merging, and the order of custom-split documents is lost. Cause: No deduplication rule configured based on document ID association; only deduplicating by text similarity causes similar fields of different documents in the same batch to be mistakenly deleted.
- Phenomenon: 401 status code is returned when calling the vector model. Cause: The vector model API key is not correctly bound in the FastGPT configuration, or the key permissions do not cover the embedding interface.
- Phenomenon: Custom-added vector model requests are identified as LLM calls. Cause: The channel is not explicitly marked as a vector model dedicated channel in channel configuration, causing requests to be routed to the LLM service.
How to Verify Proper Configuration
- Upload associated documents for a single stamping production batch, check that the index chunk length matches the preset range with no abnormal truncation.
- Submit duplicate mold inspection checklist entries, verify that the number of retained document chunks after index deduplication matches expectations.
- Call the vector embedding interface, confirm that the returned embedding results have no empty fields or format errors.
- Configure a cross-review retrieval task, check that the sorting of retrieval results matches the document relevance.
The values provided are common starting points and should be measured against the reader's own samples.