What the data for this category looks like
Data sources for this category include design drawings, material quality certificates, welding equipment parameters, welding material warehouse ledgers, non-destructive testing system records, welder qualification ledgers, and similar materials. Daily generated documents cover welding procedure qualifications, welding procedure specifications, welding material issue records, welder qualification records, weld inspection reports, and repair orders. Data is updated after each weld, batch of materials, or shift of work is completed. A single data entry includes fields such as weld number, welding material grade, equipment operating parameters, and test results. Most fields are linked to process requirements and quality acceptance standards.
What constraints do these characteristics impose on the vector models and indexes link
Data updates occur frequently and are tied to individual welds or material batches. The index link must support incremental synchronization to avoid time delays caused by full reconstruction. Multiple document types are stored together, so differentiated chunking rules must be set by document type. This preserves semantic associations between process parameters and inspection records. Many fields are linked to rigid process requirements. Indexing must retain the association logic between fields to avoid breaks in professional semantics. Weld-related data has a high proportion of industrial professional terminology. A vector model adapted for industrial scenarios must be used to ensure accurate vector representation of professional semantics.
How to set the configuration
These configuration settings are adapted for FastGPT v4.17.0.
| Config Item | Suggested Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | The single-paragraph content of welding process documents is moderately sized, avoiding semantic splicing across multiple welds |
| Overlap Length | 100–150 characters | Weld process steps have continuity. Overlapping preserves contextual association to ensure logical completeness during Q&A |
| Similarity Threshold | 0.72–0.85 | Industrial scenarios have high requirements for professional semantic matching accuracy. This threshold filters invalid results with low matching degrees |
| Single Search Limit | Top 8 entries | Welding maintenance Q&A focuses on core process parameters. Too many results will interfere with on-site decision-making |
| Re-ranking | Enabled | Original recall results may have deviations in professional term matching. Re-ranking optimizes result priority |
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 mistakes
- Phenomenon: When configuring Tongyi multi-modal vector model, OneAPI returns a call error, prompting incompatibility with OpenAI call format. Cause: The interface path and request header for model calls were not modified, and configuration parameters for the general OpenAI vector model were directly applied.
- Phenomenon: After enabling the re-ranking configuration, the online recall test does not take effect. Returned results are not sorted by re-ranking priority. Cause: The re-ranking model's call key was not bound during the knowledge base indexing link, or the re-ranking trigger switch during the recall phase was not enabled.
- Phenomenon: Incremental indexing tasks take more than 600 seconds. Logs show repeated indexing of the same batch of weld data. Cause: The unique identifier for incremental updates was not set based on weld numbers, so the system cannot identify already indexed data and repeats the indexing process.
How to confirm the configuration is properly set
- Perform a test for equipment inspection or maintenance Q&A for a single weld. Check whether the returned results include the corresponding welding process parameters, welding material requirements, or equipment fault troubleshooting steps.
- View indexing task logs to confirm that incremental updates only process newly added weld data, with no repeated indexing records.
- Call the test interface to verify that the sorted result order differs from the original recall results, aligning with the priority of professional semantic matching.
- Check vector model call logs to confirm there are no connection errors, and the vector dimensions returned by the model match the configured requirements.
The values provided are common starting points and should be measured against the reader's own samples.