What the data for this category looks like
This category covers end-to-end quality documentation for manufacturing enterprise departments including IQC and SQE. Document types include inspection specifications, sampling plans, incoming material reports, supplier audit reports, abnormality notification forms, 8D reports, and concession acceptance approval forms. Data sources cover purchase orders, drawing specifications, supplier quality certificates, data collected by inspection equipment, historical quality records, and supplier portal systems. Update frequency fluctuates with business scenarios: new incoming material reports and inspection records are generated for each batch of incoming materials, supplier evaluation and updated audit reports are completed monthly or quarterly, and temporary documents are created through irregular special audits. Documents include fields such as batch identification, supplier code, inspection parameter thresholds, and rectification closed-loop status. Two formats are supported: structured forms and unstructured long texts.
Constraints on vector models and indexing workflows
The data characteristics of this category create multiple constraints for the vector model and indexing workflow. Dispersed multi-source data requires indexes to support unified aggregation of cross-system data and metadata association. Fluctuating update frequencies require incremental indexing mechanisms to avoid resource waste and delays from full reindexing. Mixed document formats require chunking strategies that balance the integrity of short fields and the semantic coherence of long texts. Strict business judgment rules require retrieval results to accurately match dimensions such as supplier and batch, while balancing false positive and false negative risks. This prevents defective materials from entering production or shutdowns caused by incorrect rejections.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic integrity of long texts such as incoming material reports and 8D reports, while supporting chunking efficiency for short documents such as sampling plans |
| Overlap Length | 100–200 characters | Avoids semantic breaks after long text chunking, covering cross-block batch and supplier association information |
| Similarity Threshold | 0.75–0.85 | Balances the strict judgment requirements of incoming material inspection, reducing the probability of false positives and false negatives |
| Single Search Limit | Top 10 results | Ensures retrieval results cover sufficient historical quality data, while controlling return volume to avoid resource waste |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | Calibrated based on actual measurements | Matches the carrying capacity of locally deployed vector model services, avoiding exceeding concurrent request limits |
This configuration applies to FastGPT v4.17.0.
Three Common Misconfigurations
- Phenomenon: Estimated knowledge base capacity deviates significantly from actual storage usage. Cause: Chunking rules are not configured for mixed document structures. All documents are calculated using the average single-document length, leading to unreasonable capacity planning.
- Phenomenon: A
503status code error appears in the interface, and vector model requests fail. Cause: Concurrency control parameters are not adjusted fordefaultToken. Batch retrieval requests from high-frequency incoming material inspections exceed the service carrying limit. - Phenomenon: Retrieval results cannot be filtered by supplier or batch. Cause: Metadata fields such as supplier code and batch identification are not associated in the index, leading to missing retrieval dimensions.
How to Verify Successful Configuration
- Upload a single 8D report and sampling plan. Check the chunked text fragments to confirm no key business information is broken.
- Initiate a retrieval request with a specified supplier code. Verify returned results only include historical audit and inspection documents for the corresponding supplier.
- Run a batch retrieval test. Confirm no
503status code error is triggered, and the service runs stably. - Adjust the similarity threshold. Verify retrieval result recall and filtering align with incoming material inspection business judgment logic.
The values provided are common starting points, and should be evaluated against the reader's own samples.