What the Data for This Category Looks Like
Daily documents for this category include control plans, inspection records, SPC reports, finished product inspection reports, nonconforming product reviews, release orders, quality monthly reports, and more. Data sources cover MES systems, testing equipment, laboratory systems, process parameter libraries, customer standard libraries, and historical defect libraries. Data is continuously collected alongside production batches and team shifts; some static standards such as customer requirements are updated per project. Document structures include structured parameter fields and unstructured review content. Fields include inspection items, tolerance ranges, test values, review conclusions, and more, with units corresponding to individual inspection items.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Multi-source heterogeneous data sources and real-time update rhythms require retrieval and recall to support synchronized indexing and incremental updates for cross-system data. Diverse document types include both structured SPC reports and release orders, as well as unstructured review records. This requires adapted text chunking and retrieval logic for different formats. Bid document and quotation preparation requires precise matching of customer standards and historical release data, so recall results must strictly align with business scenarios to avoid interference from irrelevant content. At the same time, data permissions must be strictly isolated, only allowing calls to collection data related to in-process quality and finished product release, preventing cross-category data from being mixed in.
How to Set the Configuration
Configuration references for FastGPT v4.17.0 are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match precise quality standards and historical release data, avoid irrelevant content from being included in bid documents |
limit | Top 5–7 entries | Control the total token consumption of recall results, adapt to the concise reference needs of bid documents |
searchMode | mixedRecall | Adapt to hybrid retrieval scenarios for structured parameters and unstructured review records |
usingReRank | Enabled | Re-rank multi-source recall results, prioritize returning content most relevant to inspection standards and release rules |
datasetSearchUsingExtensionQuery | Enabled | Expand relevant quality keywords based on user bid requirements, improve recall coverage |
collectionIds | Specify the exclusive collection ID for in-process quality and finished product release | Isolate data from other categories, ensure only quality documents from this category are retrieved |
Three Common Mistakes
- Symptom: A 400 error is returned when calling the API to delete data from a specified knowledge base collection. Cause: The correct
collectionIdsparameter was not specified in the request, or the permission identifier for the corresponding knowledge base was not included. - Symptom: No recall effect occurs after using HTTP response data as a knowledge base reference. Cause: The response data was not organized into structured entries containing
title,content, andsource, which does not meet the import format requirements of FastGPT. - Symptom: No results are returned for the first retrieval of quality documents, but normal recall works on the second retrieval. Cause: Incrementally updated vector indexes were not built in real time, so the latest inspection records or release order data was not covered during the first retrieval.
How to Confirm the Configuration Is Correct
- Run a retrieval test for a single quality document, verify that the matching score of the recall results falls within the configured
similarityrange. - Check the number of results returned by the retrieval, confirm that it matches the setting of the
limitparameter. - Import a batch of historical nonconforming product review records, test whether retrieval can accurately match relevant control plan content.
- Check the multi-knowledge base retrieval configuration, confirm that the exclusive collection for in-process quality and finished product release is prioritized for retrieval.
The values provided are common starting points and should be measured against the reader's own samples.