What Data for This Category Looks Like
This category’s data primarily comes from purchase orders, drawings and specification documents, supplier quality certificates, inspection equipment logs, historical quality records, and supplier portals. Daily generated documents include inspection specifications, sampling plans, incoming material inspection reports, supplier audit reports, exception contact sheets, 8D correction reports, and concession acceptance approval forms. Data updates align with each batch of incoming material completion, monthly or quarterly supplier evaluations, and irregular special audits. Document fields include batch identifiers, sampling rules, qualified judgment items, supplier codes, and correction nodes. No additional non-business-necessary statistical value fields are included.
How These Characteristics Create Constraints for Knowledge Base Retrieval and Recall
Data sources for this category are scattered, and update frequency fluctuates with incoming material batches and supplier evaluation cycles. Retrieval must support quick association of multiple data types. Daily documents have mixed formats, including structured sampling rules and unstructured correction reports. Retrieval must adapt to recall accuracy requirements for different content types. Data includes strongly associated fields such as batch identifiers and supplier codes. Retrieval must support targeted recall by specific dimensions. Unqualified judgments and correction nodes have strong timeliness. Recall results must match currently valid versions of standard documents. External supplier-related data has confidentiality restrictions. Retrieval scope must be strictly limited to authorized datasets.
How to Configure Settings
All configurations below are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Match the professional terminology precision of incoming material inspection specifications, avoid mistakenly recalling irrelevant supplier historical records |
limit | Top 6–10 entries | Cover core reference content such as sampling and judgment for a single batch of incoming materials, control the volume of retrieval results |
searchMode | embedding | Adapt to the semantic matching requirements of professional terminology and structured sampling rules for this category |
usingReRank | Enabled | Reorder preliminary recall results by business relevance, correct deviations from semantic retrieval |
datasetSearchUsingExtensionQuery | Disabled | Avoid introducing irrelevant content unrelated to supplier quality via extended queries, ensure retrieval focuses on the business scope |
collectionIds | Specify according to the supplier dataset associated with the current batch | Only retrieve authorized knowledge bases, ensure data confidentiality |
Three Common Configuration Errors
- Retrieval results include documents from unauthorized suppliers. This occurs when
collectionIdsis not configured, and the retrieval scope covers all public datasets. This violates confidentiality requirements for supplier data. - No noticeable change in retrieval result order after enabling the re-ranking model. This happens when the
similaritythreshold is set too high. Preliminary recall results are already highly matched to the query, so re-ranking produces no observable adjustments. - The system automatically splits key business paragraphs containing batch numbers and sampling rules. This occurs when no split prohibition markers are configured for this type of document. The default word segmentation logic breaks the integrity of the fields.
How to Confirm Proper Configuration
- Run a query for a single batch of incoming materials. Verify that the
collectionIdsin retrieval results match the supplier dataset associated with the current batch. - Compare retrieval results before and after adjusting the
similaritythreshold. Confirm the re-ranking model produces observable changes to result order. - Input a test document with complete business fields. Confirm the system does not automatically split key paragraphs.
- Test only fuzzy search mode. Confirm retrieval results only return knowledge base matches, and do not call the large model to generate supplementary content.
The values provided are common starting points and should be evaluated against the reader’s own samples.