What the data for this category looks like
The typical data subjects for this category are business documents from foundries, die casting workshops, and quality departments, including casting process cards, batching lists, heat records, mold histories, first-piece inspection records, material reports, and nonconforming product review forms. Data sources include ERP material master data, melting furnace thermometers, spectrometers, MES heat records, mold management systems, and customer drawing technical specifications. Update frequency is per heat, per batch, or daily. Fields include alloy element content values, pouring temperature, mold cycle number, heat number, etc., with units such as degrees Celsius, kilograms, and cycles.
Constraints on vector models and indexing from these data characteristics
This category has high data update frequency, and single-batch data volume increases with each heat. This requires indexes to support incremental synchronization to align with real-time business workflows. Document types include structured numerical records and semi-structured process descriptions, so vector models need to support embedding logic for multi-format fields. Key business fields such as alloy batching ratio and pouring temperature are directly linked to quality risks, so results retrieved by indexes must accurately match field association relationships to avoid invalid matches across heats or mold cycles. Additionally, nonconforming product traceability requires association of multi-source data, so indexes must support cross-document associated retrieval to cover data across melting, mold, inspection, and all other process stages.
How to set configurations
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the average text length of casting process cards and heat records, avoiding single-chunk token counts exceeding model limits |
| Overlap Length | 150–200 characters | Retains critical context across chunks, ensuring complete semantics of fields such as batching ratio and pouring temperature |
| Similarity Threshold | 0.75–0.85 | Filters low-match irrelevant heat or mold data, ensuring accuracy of quality audits |
| Single Search Limit | Top 10 results | Controls the scale of returned results, avoiding redundant data interfering with core nonconforming product traceability logic |
maxToken | 4096 token | Adapts to embedding requirements for long documents such as complete heat records, avoiding truncation of critical business information |
defaultToken | 1024 token | Matches the average token count of single-chunk text, optimizing computational efficiency of vector embedding |
Three common configuration errors
- Vector embedding tasks display empty fields: Cause: Key business fields such as alloy batching ratio and pouring temperature are not included in the embedding scope, resulting in some documents being unable to generate valid vectors.
- Index update timeout: Cause:
maxTokenis not adjusted to adapt to long-text heat records, causing the embedding process to exceed the system's default timeout limit. - Retrieval results mix casting data from different mold cycles: Cause: A reasonable similarity threshold is not set, causing low-match irrelevant mold cycle data to be included in the audit scope.
How to confirm configurations are correctly set
- Upload a complete heat record document, check whether the vector embedding task status shows completed, and whether the vector dimensions after embedding meet configuration requirements.
- Enter the query "Verify the alloy batching ratio for the current heat", check whether the returned results include key values from the batching list and the corresponding heat number.
- Upload 3 records from different heats consecutively, verify whether the index synchronizes and adds new data at the update frequency, with no duplicates or omissions.
- Check the vector model call logs, confirm that the
defaultTokenconfiguration matches the average token count of single-chunk text, and no frequent truncation prompts appear.
The values provided are common starting points and should be measured against the reader's own samples.