What data for this category looks like
Data for this category comes from purchase orders, drawing specifications, supplier quality certificates, data collected by inspection equipment, historical quality records, and supplier portals. Data updates are synced with each batch of incoming material completion, monthly or quarterly supplier evaluations, and irregular audits. Daily documents include inspection specifications, sampling plans, incoming material inspection reports, supplier audit reports, abnormal contact sheets, 8D corrective action reports, and concession acceptance approval forms. Data fields include batch identifiers, inspection parameters, supplier information, non-conformity records, and judgment results, with field units matching corresponding document items.
What constraints these characteristics impose on the model access and configuration workflow
Multiple scattered data sources require configurations that support cross-format document access, adapting to multiple types of inspection records including text and images. The high-frequency update rhythm requires configurations to support near-real-time index refreshing, to avoid ledger data lagging behind actual incoming material conditions. Diverse document types and structured fields require the model to accurately extract key parameters, while controlling output deviations to prevent quality judgment errors. Additionally, the severity of adverse consequences requires strict limitation of context and reference fragment length, to ensure key information is not omitted or truncated.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
vision | Enabled | Incoming material inspection reports often include material appearance and defect images captured by equipment, requiring the model to recognize image content, adhering to the visual access specification of FastGPT v4.17.0 |
temperature | 0.1–0.3 | Stable output of structured inspection ledger data is required, to avoid random deviations affecting quality judgment results |
maxContext | 8000–12000 token | The total length of purchase orders, historical inspection records, and supplier audit reports associated with a single batch of incoming material is relatively long, requiring sufficient context to carry all associated information |
quoteMaxToken | 2000–3000 characters | Sufficient original document fragments must be retained for ledger summary, to avoid truncation of key inspection parameters or supplier information |
AIPROXY_API_ENDPOINT | Fill in the access address of the corresponding model service | Adapt to model call paths for different deployment methods, supporting local or cloud-based model access |
AIPROXY_API_TOKEN | Fill in the access token of the corresponding service | Complete identity verification for model calls, ensuring compliance of access permissions |
Three common configuration errors
- Phenomenon: The reordering function configured during online recall testing does not take effect, and the number of returned results does not meet expectations. Cause: The reordering configuration was not enabled synchronously during the knowledge base indexing stage, and only setting it in the application stage does not take effect.
- Phenomenon: When
quoteMaxTokenis set above 3000, the large language model cannot receive complete context fragments. Cause: The limit of the model's own context window is exceeded, and themaxContextparameter was not adjusted synchronously for adaptation. - Phenomenon: Uploaded incoming material inspection images cannot be recognized by the model, and the output results do not include image content. Cause: The
visionconfiguration item was not enabled, and the model only supports text input processing.
How to confirm configuration is complete
- Upload a single incoming material inspection report containing text and image attachments, initiate a test query, and check whether the model output includes the defect description content in the image, to confirm that the
visionconfiguration is effective. - After adjusting the
quoteMaxTokenparameter, check the length of the original document fragments recalled by the knowledge base, to confirm that they are not excessively truncated. - Initiate a query containing multiple batches of incoming material data, check whether the output ledger summary covers all associated suppliers, inspection dates, judgment results and other fields, to confirm that context loading is working properly.
- Verify the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfigurations, initiate a test call, and confirm that the model returns normal response results without identity verification errors.
The values provided are common starting points and should be measured against the reader's own samples.