What data for this category consists of
Data for this category originates from purchase orders, drawing specifications, supplier quality certificates, inspection equipment records, historical quality ledgers, and supplier portals. Daily documents include inspection specifications, sampling plans, incoming material reports, supplier audit reports, 8D correction forms, concession acceptance approval forms, and similar items. Individual records often include associated fields such as batch number, supplier ID, inspection date, and nonconforming item number. Update frequency adjusts with each incoming material batch, monthly or quarterly supplier evaluations, and ad-hoc audits.
What constraints these characteristics impose on knowledge base retrieval and recall workflows
The batch association and supplier binding traits of this category’s data require retrieval to precisely match associated fields such as batch number and supplier ID. This prevents irrelevant results from cross-batch or non-target supplier sources. The variable update frequency requires the knowledge base to support incremental synchronization and temporary recall of the latest audit data. This adapts to the real-time retrieval needs of each incoming material batch. The diverse document types and field structures require adaptation to different formatted text chunking and field-level retrieval. It also ensures retrieval results can quickly link to the current incoming material’s inspection and correction workflows, avoiding interference with on-site judgment.
How to set up configurations
These configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of professional inspection terminology and incoming material parameters, to avoid low-relevance results interfering with quality judgment |
limit | Top 3–5 results | Documents in this category are mostly short text with tightly linked fields; too many results will increase the large model’s context processing load |
searchMode | mixedRecall | This category has both fixed-format professional parameters (such as sampling plan AQL values) and requires semantic matching of correction logic. Hybrid mode balances accuracy and recall coverage |
usingReRank | Enabled | This category requires filtering redundant non-target documents. Reranking improves the ranking priority of highly relevant inspection reports and audit records |
datasetSearchUsingExtensionQuery | Enabled | This category requires automatic expansion of retrieval to associated supplier historical quality data based on the current incoming material batch, to expand recall scope |
collectionIds | Bind corresponding knowledge base collections by business module | This category’s data is classified by modules such as inspection specifications and supplier audits. Binding reduces retrieval scope and improves retrieval efficiency |
Three common configuration errors
- Phenomenon: Retrieval results include internal ID fields from knowledge base paragraphs. Cause: The metadata output switch for retrieval results was not disabled, causing redundant fields to be included in the large model context.
- Phenomenon: All uploaded knowledge base documents are stored in the root directory. Cause: The parent ID parameter was not passed correctly when creating documents, or the parent ID format does not meet system requirements.
- Phenomenon: The knowledge base can find matching data, but the large model does not output retrieval results. Cause: The context length of retrieval results exceeds the large model’s supported limit, or the content splicing switch for retrieval results was not enabled.
How to confirm successful configuration
- Run a simulated retrieval for a single incoming material batch, verify that returned results include the target supplier’s inspection specifications and historical correction records.
- Review the metadata of retrieval results, confirm that redundant fields such as paragraph IDs are not included.
- Upload a batch of test documents and specify a parent ID, verify that documents are correctly categorized into the corresponding directory.
- Trigger a large model call, confirm that retrieval results are correctly spliced into the prompt and output.
The values provided here are common starting points, and should be evaluated against the reader’s own samples.