What this category of data looks like
This category of data primarily comes from purchase orders, drawing specifications, supplier quality certificates, inspection equipment logs, historical quality ledgers, and supplier portals. Common daily documents include inspection specifications, sampling plans, incoming material inspection reports, supplier audit reports, exception contact sheets, 8D corrective action reports, and concession acceptance approval forms. New records are generated with each batch of incoming materials. Overall supplier evaluations are completed monthly or quarterly, and special audits are conducted irregularly. Most documents contain fields such as batch number, supplier code, inspection parameters, and non-conformity determination. Some documents must be linked to corresponding batch purchase and quality inspection processes.
Constraints on Knowledge Base Retrieval and Recall
Decentralized data sources require integrating vector libraries across multiple sources. Retrieval must cover documents from different systems including procurement, quality inspection, and supplier portals. High-frequency updates require the retrieval system to support incremental synchronization, to avoid recalling expired inspection specifications or sampling plans. Multiple document types require configuring recall weights based on business scenarios. For example, when drafting exception reports, prioritize recalling 8D corrective action records and exception contact sheets. Field accuracy requires retrieval matching to link key identifiers such as batch number and supplier code, to avoid irrelevant results across batches or suppliers. The timeliness of incoming material inspection requires retrieval response to adapt to on-site rapid drafting needs, to avoid excessive waiting time.
Configuration Settings
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Incoming material inspection documents require matching precise inspection parameters and batch information. This interval filters low-relevance results while retaining valid historical quality inspection records |
limit | Top 8 entries | Drafting exception reports requires balancing comprehensiveness and readability. 8 results can cover corrective action plans and historical exception records for common non-conformities |
searchMode | mixedRecall | Balances semantic understanding and keyword matching, adapting to fixed terminology in inspection specifications and specific problem descriptions in exception reports |
usingReRank | Enabled | When recalling mixed document types, re-ranking can prioritize 8D reports and supplier audit results that better match the current drafting scenario |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands question descriptions during drafting. For example, expanding "incoming material non-conformity" to "incoming material non-conformity 8D corrective action report" to cover more relevant documents |
collectionIds | Specify dedicated collections for incoming material inspection and supplier quality | Avoid retrieving irrelevant production process inspection documents, and accurately locate business data for this category |
Three Common Configuration Errors
- Phenomenon: In a PostgreSQL database Docker deployment environment, knowledge base indexes remain in an incomplete state for extended periods, or vector generation failure errors appear. Cause: Batch parameters for vector generation are not adjusted for bulk historical quality inspection data, and sufficient cache space is not reserved for the host machine's memory.
- Phenomenon: Calling the retrieval plugin returns empty results, with no incoming material inspection-related documents matched. Cause:
collectionIdsis not specified as the dedicated dataset for incoming material inspection and supplier quality, so the retrieval scope covers irrelevant production documents. - Phenomenon: After using an NGINX proxy, knowledge base original document links cannot be accessed normally. Cause: The source path of knowledge base documents is not replaced with the proxy domain name, causing links to point to internal unexposed resources.
How to Verify Correct Configuration
- Manually enter a drafting requirement for incoming material non-conformity, check if retrieval results include documents such as 8D corrective action reports and exception contact sheets.
- View the knowledge base index status, confirm that indexes can complete synchronization within a reasonable time after bulk updates of historical data.
- Test the original document link after NGINX proxy, confirm that it can normally jump to the corresponding supplier quality certificate or inspection specification document.
- Adjust the
similaritythreshold, verify that low thresholds recall irrelevant results, and high thresholds filter irrelevant non-quality inspection documents.
The values provided are common starting points and should be measured against the reader's own samples.