What Data Looks Like for This Category
Typical entities in the general mechanical equipment assembly scenario are manufacturers of pumps, valves, compressors, reducers, machine tools and non-standard equipment. Daily generated documents include assembly process cards, BOMs, torque records, clearance adjustment records, commissioning reports, factory inspection sheets and packing lists. Data mainly comes from PLM/ERP systems, part inspection data, assembly tool collection equipment, test benches, MES systems and customer technical agreements. Data update frequency is once per device or per order. Documents contain clear technical parameter fields such as torque values, clearance values, coaxiality indicators and similar metrics. Strong correlations exist between different documents: for example, BOMs must correspond to assembly process cards, and commissioning reports must link to torque and clearance records.
What Constraints These Characteristics Impose on Vector Models and Indexing
The update frequency of multiple orders and multiple devices requires indexes to support incremental updates. This avoids excessive computing resource usage from full reindexing. Diverse document types and structured technical parameters require chunking strategies adapted to different document lengths. For example, assembly process cards can be chunked by paragraph. Single torque records can be chunked by single collection data points. Multiple heterogeneous data sources from different systems require unified format preprocessing before indexing. This ensures semantic consistency in vector generation. The existence of strongly correlated documents requires the indexing link to retain association metadata between documents. This allows relevant documents to be returned jointly during subsequent retrieval. For example, retrieving clearance adjustment records should simultaneously return corresponding BOM information.
How to Set Configurations
This configuration is formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic integrity of long documents such as assembly process cards and commissioning reports, while controlling the computational load of vector generation |
| Overlap Length | 100–150 characters | Prevents chunking from breaking technical parameter correlations across paragraphs, ensuring semantic coherence |
| Similarity Threshold | 0.75–0.85 | Filters low-match retrieval results, focusing on documents strongly related to assembly anomalies or process requirements |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
maxToken | 8192 | Supports vector generation for long texts containing multi-level BOMs and multiple sets of test data |
| Single Search Limit | Top 10 entries | Balances retrieval efficiency and result coverage, avoiding returning too many redundant documents |
Three Common Mistakes
- Phenomenon: No options appear in the text understanding model dropdown when creating a knowledge base. Cause: The indexing model is only added to the global model channel but not linked to the current scenario, and not bound to the vector configuration link of the target knowledge base.
- Phenomenon: Indexing takes too long. Cause: Chunk length is set to more than 1500 characters, and incremental indexing is not enabled. The single processing data volume is too large during full indexing.
- Phenomenon: Retrieval returns no results after vectorizing the dataset. Cause: The similarity threshold is set too high, filtering out results that match business semantics, or the mapping relationship between document fields and retrieval keywords is not configured correctly.
How to Confirm Configuration is Complete
- Upload a single assembly process card to the knowledge base, check that the vector generation log has no errors, and confirm that the model is functioning normally.
- Initiate a retrieval test, input "reducer assembly clearance adjustment requirements", confirm that the returned results include the corresponding process card and clearance adjustment records.
- Add a BOM document for a new order, trigger incremental indexing, and check that the indexing task status shows completed.
- Enter the knowledge base configuration page, confirm that the added ollama qwen2.5 language model and bge-m3 indexing model are in an available state.
The values provided are common starting points and should be measured against the reader's own samples.