This document is written based on the vector model and index configuration logic of FastGPT v4.17.0.
What the Data for This Category Looks Like
The data for this category mainly comes from PLM/ERP, part inspection data, assembly tool collection systems, test benches, MES, and customer technical agreements. Daily documents include assembly process cards, BOMs, torque records, gap adjustment records, commissioning reports, factory inspection sheets, and packing lists. Data is updated per single device or per order. Fields include device model, order number, torque value (unit: N·m), gap value (unit: mm), coaxiality parameter (unit: μm), and other related fields.
Constraints These Characteristics Impose on Vector Models and Indexes
Multi-source heterogeneous data sources, per-device or per-order update cadence, and fields with physical units create multiple constraints for the vector model and indexing workflow. Differences in document formats across sources require retaining field association logic during chunking to avoid semantic fragmentation. The per-device update characteristic requires configuring an incremental indexing mechanism to avoid wasted computing resources from full reindexing. Fields with units such as torque and gap require associating unit semantics before embedding to improve retrieval matching accuracy. The growing dataset scale as orders increase requires an index structure that supports dynamic scaling to maintain recall efficiency.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the semantic unit length of documents such as assembly process cards and BOMs to avoid semantic breaks within chunks |
| Overlap Length | 100–150 characters | Retains cross-chunk assembly steps and parameter association information to improve retrieval coherence |
| Similarity Threshold | 0.72–0.85 | Matches fields with physical units, filters low-relevance non-assembly documents |
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 | 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. |
| Single Search Limit | Top 10 entries | Balances recall accuracy and computing overhead, adapts to scenarios with parallel retrieval across multiple orders |
Three Common Misconfigurations
- Phenomenon: After importing BOMs and torque records, retrieval fails to associate parameters corresponding to the target assembly process. Cause: Chunking did not retain context association between BOM entries and assembly steps, only field text was embedded separately.
- Phenomenon: Retrieval of torque records returns a large number of irrelevant part machining documents. Cause: Unit semantics were not uniformly associated during the embedding phase, leading to misclassification of parameters from different scenarios as similar.
- Phenomenon: Full index reindexing takes too long, failing to complete updates for a single batch of orders on schedule. Cause: No incremental indexing mechanism was configured, and full dataset reindexing occurs on every update, consuming excessive computing resources.
How to Confirm Proper Configuration
- Import a complete assembly process card and corresponding torque records, retrieve assembly parameters for a specified device, and check if returned results include complete process steps and associated field content.
- Simulate an incremental update operation for a single device, verify that only documents related to the target order are added to the index, and that historical data is not overwritten.
- Retrieve cross-document associated content, such as the assembly process corresponding to a specific part in a BOM, check if returned results cover associated information from relevant documents, and that context length does not exceed the threshold set by
maxToken. - View the index status of the vector database, confirm that the incremental update trigger logic matches the configured update rules.
The values provided are common starting points and should be measured against the reader's own samples.