What the data for this category looks like
Daily documents for this category typically include production orders, VIN lists, assembly work instructions, tightening data, offline inspection reports, rework records, and consistency release documents. Data mainly comes from ERP, MES systems, VIN and configuration management systems, tightening tools, EOL inspection lines, and quality traceability systems. Data is generated in real time per individual vehicle. Batch data is synchronized per shift. Core configuration data is finalized when a vehicle rolls off the production line. All documents are linked to full-lifecycle data using VIN as the unique identifier. Fields include part numbers, torque parameters, inspection thresholds, release status, and other relevant fields. Field units vary by document type, with no unified fixed standard.
What constraints these characteristics impose on knowledge base retrieval and recall
Single-vehicle linked data centers on VIN. Retrieval must precisely match the VIN field. This prevents mixing of data across different vehicles. Real-time and per-shift update frequencies require the knowledge base index to support incremental updates. This avoids retrieving outdated data. For scenarios with mixed multi-type document retrieval, differentiate the weight of different document types. Examples include assembly instructions and inspection reports. This prevents non-critical content from interfering with core retrieval. Single vehicles are linked to hundreds of configuration items. The number of retrieved documents per query must match the scale of a single vehicle’s linked items. This avoids retrieving redundant content or missing critical information. High-precision fields such as torque and inspection parameters require the retrieval similarity threshold to adapt to industrial documentation rigor. This prevents irrelevant content from being included in results.
How to Set Configurations
The following configurations are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapts to the rigor of industrial documentation, prevents retrieval of unrelated assembly instructions or inspection data |
limit | Top 10 entries | Matches the retrieval scale of hundreds of linked configuration items per single vehicle, controls context length |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Performs secondary sorting of initial retrieval results by relevance, corrects bias from single-vector retrieval |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms related to VIN, covers linked fields such as part numbers and configuration items |
collectionIds | Configured by workshop/document type grouping | Differentiates collections such as assembly instructions and quality inspection documents from final assembly workshops, narrows retrieval scope |
Three Common Misconfigurations
- Only preset specified responses are returned during conversations, without calling knowledge base content. The cause is incorrect configuration of tool call trigger logic, or the
limitparameter for knowledge base retrieval is set too low, leading to failure to retrieve critical documents. - Subsequent step content from the knowledge base is omitted in AI responses. The cause is the
similaritythreshold being set too high, filtering out slightly relevant but complete document fragments, or thelimitparameter being set too small, failing to retrieve all linked documents. - Reference sources are not displayed on published knowledge base pages. The cause is that the display configuration for retrieval result sources is not enabled, or document metadata fields are not correctly linked.
How to Confirm Configurations Are Correct
- Enter a single vehicle’s VIN to perform retrieval, check whether retrieved documents include assembly instructions, tightening data, and other content linked to that VIN.
- Adjust the
similarityparameter, observe changes in retrieval result relevance, confirm the threshold adapts to the rigor requirements of current documents. - Trigger the tool call workflow, verify that knowledge base results are returned when matching content exists in the knowledge base, and preset specified responses are returned when no matches are found.
- Check the reference area in the conversation interface, confirm that document sources of retrieval results are displayed normally.
The values provided are common starting points, and should be evaluated against the reader’s own samples.