What Data for This Category Looks Like
Data for this category originates from purchase orders, drawings and specifications, supplier quality certificates, inspection equipment, historical quality data, and supplier portals. Daily generated documents include inspection specifications, sampling plans, incoming material reports, supplier audit reports, exception contact sheets, 8D reports, and concession approval documents. Data updates align with each incoming material batch, monthly or quarterly supplier evaluations, and ad-hoc audits. Document fields include supplier number, incoming material batch number, inspection items, qualified judgment parameters, rectification deadlines, and some documents include physical units such as pieces, batches, and work hours.
How This Category's Data Characteristics Impact Knowledge Base Retrieval and Recall
This category's data characteristics create multiple constraints for retrieval and recall. Data sources are dispersed, and updates occur per batch, scheduled cycle, or ad-hoc audit. Retrieval must sync the latest supplier quality certificates, incoming material reports, and audit results in real time. Documents include both structured reports and unstructured content, so retrieval logic must adapt to varied text structures. It is also necessary to associate exclusive fields such as supplier number and incoming material batch number to avoid mixing results across batches or suppliers. Business scenarios require balancing risk control and production supply assurance, so retrieval and recall must balance accuracy and recall scope. This prevents false recall of non-qualified data or missed recall of critical compliance information.
How to Configure
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets compliance judgment requirements for incoming material inspection, balances accuracy and recall coverage, and avoids false recall of non-matching inspection specifications |
limit | Top 8 entries | Supports scenarios where a single audit needs to reference multiple sampling plans and historical reports, avoids excessive results increasing manual verification workload |
searchMode | mixedRecall | Balances structured field matching and unstructured text semantic recall, compatible with FastGPT v4.17.0 and later hybrid retrieval capabilities |
usingReRank | Enabled | Re-ranks recall results, prioritizes documents with the latest audit date and matching batch numbers, improving visibility of critical information |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands queries to include associated terms for suppliers, batches, and inspection types, adapting to this category's business-dimensional retrieval needs |
collectionIds | Collection of supplier quality knowledge base IDs divided by business domain | Limits retrieval scope to documents exclusive to this category, avoiding mixing irrelevant data from other categories |
Three Common Configuration Mistakes
- Issue: The error
invalid configuration parameter name "hnsw.max_scan_tupleappears during retrieval. Cause: Non-specified configuration items were used incorrectly, with external vector database parameters mistakenly added to FastGPT retrieval configurations. - Issue: Single retrieval takes excessive time, causing interface loading lag. Cause: No reasonable
limitparameter set, leading to too many irrelevant recalled documents, or nocollectionIdslimit applied, resulting in an overly large retrieval scope. TheusingReRankoptimization logic was also not enabled. - Issue: Retrieval and recall rate for scanned PDF documents is extremely low, with no valid business content extracted. Cause: No OCR preprocessing was configured for image-based PDFs. Only plain text PDFs were chunked and embedded, so inspection data and audit conclusions in scanned content cannot be recognized.
How to Verify Successful Configuration
- Review the configuration item list to confirm only specified retrieval parameters are used, with no unauthorized additional parameters.
- Run a query that includes a supplier number and incoming material batch number, verify that recall results are limited to the knowledge base for the corresponding business domain.
- Upload a scanned PDF test document, confirm the retrieval process can extract text content from images.
- Test batch query scenarios with multiple questions, confirm the retrieval interface returns matching results grouped by batch.
The values provided are common starting points and should be evaluated against the reader's own samples.