What the Data for This Category Looks Like
Data sources include PLM/ERP systems, part inspection data, assembly tool collection records, test bench data, MES systems, and customer technical agreements. Daily documents include assembly process cards, BOMs, torque records, clearance adjustment records, commissioning reports, factory inspection sheets, and packing lists. Fields include material codes, torque values (unit N·m), coaxiality values (unit mm), tightening time, and others. Documents are updated once per device or per order.
Constraints Imposed on Vector Models and Indexes
This category’s multi-variety assembly per order trait requires indexes to split collections by order dimension, to prevent mixing document chunks from different orders. High-frequency single-device updates require indexes to support incremental updates, cutting full rebuild time. Multiple structured fields with units require vector models to adapt to structured metadata encoding, improving association matching accuracy. Multi-source heterogeneous data formats require preprocessing before indexing to unify field names and unit specifications, reducing subsequent association errors.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapt to the information density of single segments in documents such as assembly process cards and BOMs, avoid scattered information within chunks |
| Overlap Length | 50–80 characters | Retain contextual association for segmented content such as torque records and clearance adjustment records |
| Similarity Threshold | 0.75–0.85 | Match the association degree between assembly processes and physical inspection data, filter low-correlation content |
| Single Search Limit | Top 10 entries | Adapt to the scale of document associations for single-device assembly, avoid redundant information interfering with reviews |
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. |
| Maximum Context | 3500 token | Cover the length of associated information for full-process documents of a single device |
Three Common Misconfigurations
- Phenomenon: The interface displays "request timed out" or returns a 504 status code when switching knowledge base indexes. Cause: Index collections are not split by order dimension, and single-batch index data volume exceeds system processing thresholds.
- Phenomenon: Knowledge base content is limited but remains in "rebuilding" status for more than 24 hours. Cause: Incremental index updates are not enabled, and each rebuild triggers full processing of all assembly documents. This category’s documents are updated frequently per single device, so full rebuilds take excessive time.
- Phenomenon: The interface returns a "missing required fields" error when adding indexes in bulk. Cause: Chunk metadata is not configured according to the category’s dedicated document field structure, and mappings for key fields such as torque values and coaxiality are omitted.
How to Confirm Proper Configuration
- Manually upload an assembly process card and its corresponding torque record, check the chunk association results after index generation, confirm similarity matching meets the expected threshold.
- Initiate a small-batch index rebuild for the document volume of a single device, check system time consumption and status, ensure no timeout errors or abnormal stalls.
- Check the metadata fields of the index collection, confirm that category-specific information such as BOM material codes and torque value units is included.
- Call the bulk index addition test interface, pass sample data conforming to the category’s document structure, confirm the interface returns successfully and metadata is complete.
The values listed are common starting points, and should be evaluated against the samples relevant to the specific deployment.