What the Data for This Category Looks Like
Data sources for heat treatment scenarios include material grades, customer technical requirements, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers, metallographic equipment, and manufacturing execution systems (MES). Daily documents cover heat treatment process cards, furnace loading diagrams, furnace temperature curves, hardness and metallographic reports, equipment inspection records, and exception handling sheets. Explicit fields include furnace temperature (℃), hardness (HV), and metallographic structure type, among others.
Data is updated per furnace charge and per batch. A single batch of data includes associations between multiple document types. Individual document lengths vary widely, ranging from short-text inspection records to lengthy furnace temperature curve analysis reports.
What Constraints Do These Characteristics Create for Vector Models and Indexing
The high-frequency updates and mixed multi-type data characteristics of heat treatment scenarios create multiple constraints for vector models and indexing workflows.
High-frequency updates per furnace charge require indexes to support low-latency incremental writes, avoiding additional time costs from full reconstruction. Mixed storage of multi-type documents requires chunking logic to adapt to different structures of text, tables, and text embedded in images, preventing loss of key information. Explicit fields and units require indexes to retain metadata associations for data, ensuring retrieval results can directly match specific process parameters and test data. Cross-document association requirements require indexes to support joint retrieval across process cards, furnace temperature curves, and metallographic reports, enabling full-link traceability for quality anomalies.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph text length of heat treatment process cards and furnace temperature curves, avoiding cutting off process logic and temperature change nodes |
| Overlap Length | 150–200 characters | Retains time-series context for furnace temperature curves, preventing chunking from splitting continuous temperature fluctuation data |
| Similarity Threshold | 0.72–0.85 | Distinguishes normal process fluctuations from abnormal structural deviations, matching the structural similarity judgment standards in metallographic reports |
maxToken | 8192 | Adapts to the input upper limit of multimodal embeddings, covering the total token count of process card text and furnace temperature curve image descriptions |
| Single Search Limit | Top 6 results | Matches the number of documents associated with each furnace charge, avoiding redundant retrieval results that interfere with analysis |
Three Common Configuration Errors
- Setting the maximum chunk paragraph depth to 3 will return empty results when retrieving annotated text from furnace loading diagrams. This occurs because the model's automatic paragraph recognition function is not enabled, and chunks are only split by physical line breaks, without extracting annotated text embedded in charts.
- An Invalid error is returned when accessing a multimodal embedding model. This occurs because the file types for multimodal input are not configured correctly. The image format of heat treatment furnace temperature curves is not on the model's supported list, or the request parameters do not include the correct file fields.
- Retrieval results for old vector database data deviate after upgrading the version. This occurs because the original index's field mapping rules are not retained, and the new configuration's vector dimensions do not match the old database, causing similarity calculation failures.
How to Verify the Configuration Is Correct
- Upload a single heat treatment process card, and check if the chunked text fragments fully retain process steps and parameter associations.
- Submit a document containing a furnace temperature curve image, verify that the multimodal embedding returns no errors, and that abnormal curve nodes can be matched during retrieval.
- Perform a batch index reconstruction operation, check the progress status of the task queue, and confirm that the index status of all documents is complete.
- Compare retrieval results from the old and new vector databases, and confirm that migrated documents can return association results consistent with the original database.
The values provided are common starting points and should be measured against the reader's own samples.