What this category of data looks like
Data for the general mechanical equipment assembly category comes primarily from PLM/ERP, part inspection data, assembly tool collection systems, test benches, MES, and customer technical agreements. Updates occur once per device or per order. Daily documents include assembly process cards, BOM lists, torque records, clearance adjustment records, commissioning reports, factory inspection certificates, and packing lists. Most data fields include specific numerical parameters such as torque values, clearance dimensions, and coaxiality indicators, with clear unit annotations attached.
Constraints on vector models and indexing
Data for this category is updated per device or order, and includes many process parameters with units and step-by-step text. This creates multiple constraints for the vector model and indexing workflow. First, incremental data per device or order must support fast index updates. This avoids the time cost of full reindexing. Second, numerical parameters with units must retain associated context during chunking. This prevents loss of the binding relationship between parameters and units after text is split. Third, document types vary widely with significant structural differences. The index must support mixed embedding of structured parameters and unstructured process text.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-section text length of assembly process cards and BOMs, avoids losing context association of process steps after splitting |
| Overlap Length | 100–150 characters | Retains parameter and step connection information for adjacent chunks, prevents text with continuous numerical values such as clearance adjustment records from being split apart |
| Similarity Threshold | 0.72–0.80 | Filters low-match retrieval results, aligns with the requirement for parameter precision in equipment inspection |
| Maximum Context | 6000 characters | Matches the context length required for process Q&A in general mechanical equipment assembly scenarios, avoids exceeding model input limits |
defaultToken | Set based on actual measurement | Adapts to the token limit of different embedding models, supports embedding of numerical text with units such as torque and clearance |
maxToken | 8192 | Matches the maximum token limit of mainstream open-source embedding models, ensures long process text for assembly work can be fully embedded |
Three Common Misconfigurations
- Symptom: The default index inserted manually into the knowledge base disappears automatically several hours later. Cause: Persistent storage for incremental indexing is not configured, or the index refresh cycle setting does not match the data update frequency. Temporary indexes are automatically cleaned up in this case.
- Symptom: After connecting qwen3-embedding-8b, the file status continuously displays "Indexing". Cause: The interface timeout setting of the selected embedding model does not adapt to the loading time of the large model, or compatible configuration between the model version and FastGPT v4.17.0 is not completed.
- Symptom: Retrieval results show mismatched content where parameters and units are separated. Cause: Chunk length is set too short, causing numerical parameters with units to be split into different chunks. Their associated relationship is lost after embedding.
How to Confirm Configuration is Complete
- Upload a test document containing torque records and clearance adjustment records. After indexing is complete, search for "how to adjust reducer clearance". Check if the returned results include the corresponding parameters and steps.
- Upload a new order document to trigger incremental indexing. Check if the indexing status completes within a reasonable time, and no continuous "Indexing" status appears.
- After replacing the embedding model, perform a batch re-embedding operation. Check if the retrieval recall rate of existing documents meets expectations.
- View the index storage directory. Confirm that index files are not automatically cleaned up, and the persistent configuration is active.
The values provided here are common starting points, and should be evaluated against the reader's own samples.