What the Data for This Category Looks Like
Data sources include customer drawings, special characteristic lists, ERP/MES systems, measurement and inspection tool data, equipment process parameters, and supplier quality certificates. Daily generated documents include control plans, process flowcharts, work instructions, first article and patrol inspection records, SPC reports, PPAP materials, and 8D reports. Data is collected synchronously per production batch and each shift. Some equipment parameters update continuously in real time. Document structures cover structured parameter tables, process steps, inspection values, and quality records. Fields include special characteristic numbers, process parameters, inspection thresholds, and other relevant fields. Common industrial units such as millimeters, newtons, and megapascals are used.
Constraints for Knowledge Base Retrieval and Recall
Dispersed data sources and frequent updates require the retrieval link to support fast recall of multi-source data. This avoids review deviations caused by delayed data. Diverse document structures and strongly constrained special characteristic fields require retrieval to accurately match core parameters. Broad recall of irrelevant content is not allowed. Severe consequences of errors in mass production require recall results to cover all process-related documents. This prevents missing key quality records. Constraints from vehicle customers require strict limitation of the retrieval scope to the knowledge base of the corresponding business domain. Cross-category irrelevant data must not be included.
Configuration Settings
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.85–0.92 | Match core fields such as special characteristics and process parameters of automotive parts, to prevent low-relevance results from entering the cross-review process |
limit | Top 6–10 results | Control the context window length, to meet the content capacity needs of multi-document cross-review in mass production scenarios |
usingReRank | Enabled | Run secondary ranking on initial recall results, filter out miscellaneous documents with mismatched formats, and improve review efficiency |
datasetSearchUsingExtensionQuery | Enabled | Generate extended queries based on the original question, to cover retrieval needs for specific fields such as customer drawing numbers and PPAP numbers |
collectionIds | Bind exclusive collections by business domain | Only retrieve knowledge bases of corresponding business domains such as braking and chassis, to avoid interference from cross-category irrelevant data during reviews |
Three Common Configuration Mistakes
- Phenomenon: A locally deployed 14B model returns an error or empty result after a knowledge base is added. Cause: The association binding between the knowledge base and the large model is not configured properly, or the
similaritythreshold is set too high, making it impossible to match the embedding vectors of local documents. - Phenomenon: The number of retrieved results is far lower than expected, with only 1–2 results returned. Cause: The
limitparameter is mistakenly set to an excessively small value, ordatasetSearchUsingExtensionQueryis not enabled, failing to cover multi-field retrieval requirements. - Phenomenon: Retrieved results include documents from non-target categories, such as work instructions for general machinery. Cause: Exclusive knowledge base collections are not bound via
collectionIds, causing the retrieval scope to cover data from irrelevant business domains.
How to Verify Successful Configuration
- Input a clear special characteristic number, and check whether the retrieved results include the corresponding control plan or work instruction document.
- Check the number of retrieved results, confirm that it falls within the setting range of the
limitparameter. - Verify the source collection of the retrieved results, confirm that only knowledge base documents from the current business domain are included.
- Trigger a simulated cross-review process, confirm that the system does not return content outside the knowledge base scope.
The values provided are common starting points and should be measured against the reader's own samples.