What this category of data looks like
The data for this category mainly comes from purchase orders, drawing specifications, supplier quality certificates, inspection equipment, historical quality data, and supplier portals. Daily documents include inspection specifications, sampling plans, incoming material reports, supplier audit reports, 8D reports, concession acceptance approval forms, and other related materials. Some documents contain dedicated fields such as version numbers, batch numbers, AQL values, and inspection results. Data updates generate new incoming material reports with each batch of incoming materials, complete supplier evaluation updates on a monthly or quarterly basis, and conduct audits to update audit reports irregularly.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The data characteristics of this category create multiple constraints for retrieval and recall. Documents from multiple sources with large structural differences require that the retrieval scope be accurately limited to the corresponding business dataset to avoid interference from irrelevant content. Dedicated fields with batch numbers and version numbers require retrieval to match business-specific identifiers, ensuring recalled results correspond to the current incoming material batch and supplier version. Frequently updated business data requires that recalled results prioritize returning the latest version of documents. Mixed structured and unstructured document types require retrieval to support content parsing for multiple formats such as tables and text, ensuring key fields can be properly indexed.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
collectionIds | Dedicated dataset IDs for incoming material inspection and supplier quality | Limit the retrieval scope to avoid mixing irrelevant documents from other categories, ensuring recalled content matches the current business scenario |
similarity | 0.75–0.85 | Balance the precision and coverage of recalls. Setting it too low introduces irrelevant content, while setting it too high causes valid and compliant documents to be missed |
limit | Top 8 entries | Control the number of returned results, adapt to the quick review needs of IQC and SQE personnel, and avoid information overload |
usingReRank | Enabled | Re-rank initial recall results, correct deviations from pure similarity sorting, and prioritize content with higher matching degree to business questions |
datasetSearchUsingExtensionQuery | Enabled | Generate extended queries based on the original question, adapt to retrieval of dedicated fields such as batch numbers and version numbers, and improve the recall effect of structured documents |
searchMode | embedding | Adapt to semantic matching requirements for multiple types of documents, and better cover business logic associations across documents |
Three common configuration errors
- Phenomenon: After calling knowledge bases corresponding to different categories in a workflow, some results have empty business fields and cannot enter the unified processing logic. Cause: A unified
collectionIdsrange is not configured for all business scenarios, and only the dataset corresponding to a single category is bound, resulting in cross-category retrieval results not being included in the processing link. - Phenomenon: The sorted results of combined knowledge base searches do not conform to business priorities, and documents with high compliance are not ranked first. Cause:
usingReRankis not enabled, and sorting only relies on the initialsimilarityvalue without adjusting priorities based on business-specific identifiers such as batch numbers and version numbers. - Phenomenon: Batch numbers, AQL values and other fields in uploaded XLSX format incoming material reports cannot be extracted. Cause: The retrieval extension configuration for structured documents is not enabled, and the table content parsing function is not turned on, resulting in non-text content not being properly indexed.
How to confirm the configuration is correct
- Check whether the
collectionIdsin the retrieval configuration includes all datasets related to incoming material inspection and supplier quality, ensuring no omissions. - Trigger a retrieval simulation for an incoming material-related question, verify that the
similarityvalues of returned results fall within the preset range, and that the number of results matches thelimitconfiguration. - Check whether the sorted retrieval results prioritize displaying the latest version of documents, verifying that
usingReRankis active. - Upload a test XLSX format inspection report, confirm that fields such as batch numbers and inspection results in the table can be properly retrieved.
The values given are common starting points and should be measured against the reader's own samples.