What data for this category looks like
Data comes from material grades, customer technical specifications, temperature control instruments, furnace temperature uniformity testing equipment, hardness testers, metallography equipment, and manufacturing execution systems (MES). Updates run per furnace load and per batch. Daily documents include heat treatment process cards, furnace loading diagrams, furnace temperature curves, hardness and metallography reports, equipment inspection forms, and exception handling sheets. Document fields cover heating temperature, holding time, cooling rate, furnace number, batch number, hardness value, metallographic structure type, and more. Some documents include visual chart data.
What constraints these characteristics impose on vector models and indexing
High-frequency updates per furnace load and per batch require indexes to support incremental refreshes, to avoid performance loss from full reindexing. Mixed document types require vector models to adapt to both structured parameter text and unstructured chart descriptions. This ensures retrieval can associate and match process parameters and exception records. Strong field associations exist within single documents. For example, furnace numbers must be bound to corresponding furnace temperature curves and hardness reports. Chunking and index design must retain field associations, to prevent losing critical matching clues after splitting. High retrieval precision is required for exception data. Ensure recall results cover all potential non-conforming items, and avoid missing key audit basis.
How to set configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-section content length of heat treatment process cards and exception handling sheets, avoids splitting that breaks associations between process parameters and exception descriptions |
| Overlap Length | 100–150 characters | Ensures sufficient field overlap between adjacent chunks, prevents key information from being split across two chunks |
| Similarity Threshold | Calibrated via actual testing | Determined by testing recall effects across different exception scenarios, aligned with precision requirements for heat treatment audits |
| Single Search Limit | Top 10 results | Matches the typical number of audit data associated with a single furnace load, avoids retrieving excessive irrelevant content |
maxToken | 8192 | Adapts to the maximum token limit of mainstream Chinese embedding models, covers the typical text length of heat treatment documents |
defaultToken | bce-embedding-zh | Adapts to semantic understanding of Chinese process documents, meets embedding needs for domestic scenarios |
Three common misconfigurations
- An "insufficient model permissions" error occurs when calling embedding models. This is caused by incorrect embedding model channel configuration, or uncompleted access permission activation for the channel.
- Knowledge base search takes longer than expected. This is caused by overly large chunk length and disabled incremental indexing, or mismatched hardware resources for batch retrieval tasks.
- Retrieval results miss key content from exception handling sheets. This is caused by splitting furnace number and exception description associated fields during chunking, leading to failure to match complete audit clues during retrieval.
How to confirm proper configuration
- Upload a single heat treatment process card, check that chunked text retains complete key parameter fields such as heating temperature and holding time.
- Initiate a search with specific furnace number and hardness value keywords, confirm that recall results include corresponding furnace temperature curve and metallography report data.
- View embedding model call logs, confirm there are no token permission-related error messages.
- Batch upload 10 heat treatment documents from the same batch, check that index update latency meets expected levels.
The values provided are common starting points, and should be evaluated against the reader’s own samples.