What the data for this category looks like
Data for the complete vehicle final assembly scenario centers on the Vehicle Identification Number (VIN) of a single vehicle. Daily covered documents include production orders, VIN lists, assembly work instructions, tightening data, offline inspection reports, and rework records. Data sources cover ERP/MES systems, VIN and configuration systems, supplier batch libraries, tightening tools, EOL inspection lines, and quality traceability systems. Update frequencies follow three rhythms: per vehicle, real-time, and per shift. Data includes structured fields such as VIN, torque value, inspection items, and part batch numbers, as well as unstructured content like assembly steps and troubleshooting guides.
What constraints these characteristics impose on knowledge base retrieval and recall
Single-vehicle data centered on VIN requires retrieval to precisely match the configuration and part information of the current vehicle model, to avoid cross-model data interference during mixed-line production. Multi-source, multi-update-frequency data requires the knowledge base to support incremental synchronization and real-time recall, to ensure retrieval results align with production tempo. The coexistence of structured fields and unstructured documents requires retrieval to support both keyword matching and semantic recall, balancing accuracy and content completeness. The need to access hundreds of associated configuration data entries per single vehicle requires controlling the reasonable scope of recall results to avoid exceeding model context limits, while covering necessary associated information.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match the precision of structured fields such as VIN and torque values, filter low-correlation cross-model configuration data |
limit | Top 10 entries | Cover the recall needs of core associated data per single vehicle, while controlling the context length of single-round retrieval |
searchMode | mixedRecall | Match both structured keywords such as VIN and semantic content, adapt to precise retrieval scenarios in mixed-line production |
usingReRank | Enabled | Reorder recall results by matching degree with the current VIN, prioritize returning model-specific work instructions and inspection standards |
datasetSearchUsingExtensionQuery | Enabled | Automatically expand retrieval for associated data such as supplier batches and tightening tool parameters based on the input VIN |
collectionIds | Bind specified knowledge base collections by vehicle model/workstation | Limit retrieval scope, avoid recalling irrelevant documents for current workstation or model, reduce retrieval noise |
Three common misconfigurations
- Phenomenon: Retrieval results only include document text, and step diagrams in assembly work instructions are not recalled. Cause: The image parsing plugin of FastGPT is not enabled, and only the text content of PDF documents is embedded and recalled.
- Phenomenon: Saved retrieval configuration parameters revert to default values after service restart. Cause: Configuration is only stored in local temporary cache, not persisted to the system database.
- Phenomenon: Retrieval by VIN returns a large number of configuration data for irrelevant vehicle models. Cause:
searchModeis not configured as hybrid retrieval mode, relying only on vector recall leading to semantic matching deviation.
How to verify correct configuration
- Upload a final assembly work instruction with assembly diagrams, enter the target VIN to initiate retrieval, confirm that the returned results include text content and corresponding image links.
- Manually modify the retrieval configuration parameters, submit and save them, re-enter the configuration page, confirm that the parameters do not revert to default values.
- Enter the VIN of a specific vehicle model to initiate retrieval, check whether the returned results include model-specific documents such as tightening data and offline inspection reports.
- Use FastGPT v4.17.0, view the system retrieval log, confirm that both vector retrieval and keyword retrieval are triggered, and the reranking function works normally.
The values given are common starting points and should be measured against the reader's own samples.