What data for this category looks like
Data for this category covers the full end-to-end process of incoming material inspection and supplier quality management for manufacturing enterprises. Data sources include purchase orders, drawing specifications, supplier quality certificates, data exported from inspection equipment, historical quality ledgers, and audit materials submitted via supplier portals. Daily generated documents include inspection specifications, sampling plans, incoming material inspection reports, supplier audit reports, exception contact sheets, 8D corrective action reports, and concession approval forms. Data update frequency aligns with business milestones: new inspection data is generated for each incoming material batch, overall supplier evaluations are updated monthly or quarterly, and special audits are conducted irregularly with corresponding documents produced. Data fields include batch numbers, supplier names, inspection parameters, defect types, corrective action statuses, and some fields are enumeration or numeric values.
What constraints do these characteristics impose on vector models and indexing
Multi-source heterogeneous data sources require indexes to support mixed-format vector embeddings, and compatibility with both structured purchase order data and unstructured inspection report documents. High-frequency update needs mean index configurations must support incremental synchronization, avoiding computational overhead from full reindexing every time. Document lengths vary widely: inspection specifications are long texts, while incoming material reports are short documents. This requires a flexible chunking strategy. Business identifiers such as batch numbers and supplier names in fields require indexes to support metadata-based filtering and recall, improving retrieval accuracy and preventing invalid cross-supplier data from appearing in results.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the average length of documents such as incoming material reports and inspection specifications, avoids overly fragmented chunks that lose context, or chunks too long to exceed model limits |
| Overlap Length | 100–150 characters | Retains associated information between adjacent chunks, prevents business identifiers such as batch numbers and supplier names from being split across chunks |
defaultToken | 4096 | Adapts to the input upper limit of most multimodal and text vector models, supports embedding requirements for image data exported from inspection equipment |
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. |
| Similarity Threshold | Calibrated based on actual testing | Must be adjusted in line with incoming material inspection defect judgment standards, prevents false recall of qualified batch data or missed recall of abnormal records |
| Single Search Limit | Top 10 results | Matches the batch data processing scale used by SQE teams for daily traceability, balances retrieval efficiency and recall coverage |
Three common misconfigurations
- Phenomenon: A call failure occurs when connecting to the Tongyi multimodal vector model, with an error returned stating the specified model is not supported. Cause: The multimodal model’s call parameters are not configured correctly, or the corresponding model’s key and endpoint are not registered on the platform.
- Phenomenon: Significant deviations appear in retrieval results after mixing different index models and text vector models. Cause: Vector embedding dimensions and similarity calculation rules are not unified, making cross-model vector spaces unable to align.
- Phenomenon: New batch incoming material inspection data cannot be retrieved quickly, and the retrieval result is empty. Cause: Incremental index update mode is not configured, and full reindexing is used for each index update, causing new data to not be added to the index in a timely manner.
How to confirm the configuration is correct
- Upload a historical incoming material inspection report, run an associated retrieval, and check if returned results include relevant records from the same supplier and batch.
- Initiate a batch retrieval using a supplier number, verify that the number of returned results matches the preset single search limit.
- Submit a new exception contact sheet, wait for the index update to finish, then run a retrieval to confirm the new data is included in the retrieval scope.
- Check the vector model call logs in FastGPT v4.17.0, confirm there are no error messages about unsupported models or exceeded parameter limits.
All configuration values listed above are common starting points. Adjustments should be made based on specific samples and business requirements for the deployment.