What Data for This Category Looks Like
This category covers daily quality documents and business data from manufacturing enterprise departments including incoming material inspection (IQC) and supplier quality engineering (SQE). Data sources include purchase orders, drawing specifications, supplier quality certificates, data collected by inspection equipment, historical quality records, and supplier portal information. Update frequency changes with each batch of incoming materials, monthly/quarterly supplier evaluations, and ad-hoc audits. Document types include inspection specifications, sampling plans, incoming material reports, supplier audit reports, exception contact sheets, 8D rectification reports, and concession acceptance approval forms. Fields include batch identifiers, inspection parameters, judgment results, supplier qualifications, rectification records, and some fields are associated with corresponding business units.
Constraints on Vector Models and Indexes
Multi-source and heterogeneous data sources require indexes to support cross-system and cross-format data integration, avoiding information silos. The high-frequency update rhythm requires adapting to incremental indexing mechanisms to reduce resource consumption from full reindexing. Diverse document structures and specific business fields require chunking strategies that balance the information integrity of long documents and the semantic coherence of short documents. Professional terminology and associated logic in quality scenarios require vector models to have semantic understanding capabilities in the corresponding domain, while indexes need to support filtered retrieval by key fields such as batch and supplier to ensure information accuracy during tender and quotation processes.
How to Configure
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapt to the information density per chunk of documents such as inspection specifications and sampling plans, avoiding fragmentation of critical business information within chunks |
| Overlap Length | 100–150 characters | Preserve key fields such as cross-chunk batch numbers, inspection items, and supplier IDs, preventing core information from being truncated |
| Similarity Threshold | 0.72–0.85, calibrated via actual testing | Distinguish semantic differences between qualified and unqualified quality descriptions, adapting to the precise matching needs of quality scenarios |
| Single Search Limit | Top 10 results | Cover multiple associated documents for the same batch of incoming materials, avoiding recall of redundant historical data |
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. |
| Reranking | Enabled | Optimize the ranking of quality-related information across documents, improving the accuracy of tender and quotation associations |
Three Common Misconfigurations
- Phenomenon: In the knowledge base creation page of FastGPT v4.17.0, no target options appear in the text understanding model dropdown menu. Cause: The embedding model and language model have not been associated and bound via the model channel, or the model deployment has not passed health checks.
- Phenomenon: The number of vector index recall results does not match expectations, and irrelevant batch quality data is mixed in. Cause: Filter rules based on incoming material batch and supplier ID have not been configured, causing the retrieval scope to cover all historical data.
- Phenomenon: Bulk vector model retraining cannot be triggered, and only manual parameter adjustment and upload per file are supported. Cause: The incremental update configuration for the knowledge base has not been enabled, or the trigger threshold for bulk tasks has not been set. The system defaults to supporting single-file processing only.
How to Confirm Proper Configuration
- Upload a standard incoming material inspection report, check the vector import log to confirm that the chunk length matches the set value.
- Enter a search term containing a specific supplier ID, check whether the recall results include the corresponding supplier’s audit reports and rectification records.
- Test batch upload of 3 to 5 historical quality documents, confirm that the indexing task completes automatically with no error prompts.
- Enter the model management page of FastGPT v4.17.0, confirm that all configured embedding models and language models are shown as available.
The values provided here are common starting points and should be evaluated against the reader’s own samples.