What the Data for This Category Looks Like
Typical data sources include material grade archives, customer technical requirements, temperature control instrument logs, furnace temperature uniformity test reports, hardness tester and metallographic equipment test results, and workshop MES systems. Daily documents include heat treatment process cards, furnace loading layout drawings, real-time furnace temperature curves, hardness and metallographic analysis reports, equipment inspection forms, and exception handling forms. Data is updated per furnace load and per batch. Individual documents include specialized fields with clear units such as furnace temperature (℃), holding duration (hours), and hardness value (HV). Some documents include metallographic structure images.
Constraints These Characteristics Impose on Vector Models and Indexes
High-frequency updates per furnace load and per batch require incremental indexing to avoid redundancy from full reconstruction. Documents include structured tables, unstructured text, and images. This requires support for both text and multimodal vector indexing. Specialized fields with clear units require vector models to align unit-related semantics, preventing parameter confusion. Exception handling forms and process cards have high specialized terminology density. Proper chunk length control is needed to retain contextual association, while adapting to high-frequency updated data volumes.
Configuration Setup
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the length of single-segment specialized content in heat treatment process cards and furnace temperature curves, avoiding splitting critical parameters across chunks |
| Overlap Length | 100–150 characters | Retains temporal association of furnace temperature curves across chunks, preventing chunking from disrupting process logic |
| Similarity Threshold | 0.75–0.85 | Filters historical furnace load data unrelated to queries, improving retrieval accuracy |
| Maximum Context | 4000 tokens | Covers core parameters of a complete single process card, meeting context requirements for model inference |
defaultToken | 1024 | Adapts to the default input length of the bge-m3 model, balancing indexing efficiency and semantic completeness |
maxToken | 8192 | Matches the maximum context support range of ollama qwen2.5, avoiding truncation of critical content |
Three Common Misconfigurations
- Phenomenon: When creating a knowledge base, the text understanding model dropdown does not display added ollama qwen2.5 or bge-m3 models. Cause: The local call address or port mapping for the corresponding model was not configured on the model management page, so the platform cannot recognize available models.
- Phenomenon: After uploading heat treatment documents, the system initially shows 8 segments, then the status changes to 13 segments with duplicate shards. Cause: Incremental indexing failed to properly clean up old version shards, or duplicate verification failed during upload, leading to repeated generation of vector shards.
- Phenomenon: Only single-file parameter adjustment and re-upload are supported. Batch triggering of vector model training is not available. Cause: The configuration switch for batch index training was not enabled in system settings, so the platform only provides single-file training access.
How to Confirm Proper Configuration
- Upload a single standard heat treatment process card, confirm the number of chunks matches expectations, with no critical parameters split across chunks.
- Submit a query containing furnace temperature parameters, confirm the similarity of returned results falls within the set threshold range.
- Batch upload 3 to 5 pieces of same-type furnace load data, confirm no duplicate shards exist after index update, and the number of chunks matches expectations.
- Enter the model management page, confirm both embedding models and language models are loaded normally, with no error prompts.
The values provided are common starting points and should be measured against the reader's own samples.