What data for this category looks like
Steel trade data primarily comes from spot quotation systems, steel mill production ledgers, logistics waybills, and customer inquiry records. Structured data includes fields such as steel grade, specification, origin, pricing unit (ton/meter), current inventory, and price per ton. Non-structured data mostly consists of contract scans and offline quotation sheet images. Spot quotation data is updated daily. Production ledgers are updated weekly or monthly. Logistics waybills are synced in real time with transactions. Document formats include structured tables, PDF scans, and chat log screenshots. Some documents contain multi-page, complexly formatted quotation details.
How these characteristics impose constraints on model integration and configuration
The coexistence of structured and unstructured data requires model integration to support both text and image parsing. OCR functionality must be configured to adapt to scanned document files. High-frequency updated data sources require the knowledge base sync interval to match the data update rhythm. This prevents recalling outdated quotation or inventory data. Multi-field structured data requires the model to accurately identify and match field units. This avoids pricing unit errors. Marketing content generation must combine real-time inventory and quotations. Thus, the knowledge base recall range must be limited to steel trade-related categories. This prevents the introduction of irrelevant data.
Recommended configuration values
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Steel trade documents often contain multi-page scans and multiple structured tables. Parsing takes longer than general documents |
UPLOAD_FILE_MAX_SIZE | 800 MB | Bulk steel quotation sheet scans have large individual file sizes. This setting adapts to large file upload requirements |
maxContext | 8000–12000 characters | Multiple structured quotations and unstructured contract data must be recalled simultaneously. The context length must cover multiple sets of business fields |
enable_ocr | Enabled | Offline quotation sheets and waybills for steel trade are mostly in image format. OCR is required to extract text content |
recall_top_k | Top 6 entries | Steel trade customer inquiries often focus on 1 to 2 steel grades. Too many recalled entries will interfere with the accuracy of model outputs |
similarity_threshold | 0.75–0.85 | Structured field matching requires high precision. This filters low-relevance non-target category data |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis. Testing on local samples is recommended before finalizing settings.
Three common configuration errors
- Calling a third-party model via OneAPI returns a 400 error, with a prompt indicating invalid parameters. Cause: The
modelparameter is not configured to exactly match the model identifier in the OneAPI backend, or unsupported request parameters are passed. - Updating config.json does not update the model list. Checking the container shows no changes to files in the /app/data/config directory. Cause: The local config.json is not mounted to the corresponding container directory, or the FastGPT container is not restarted after modifying the configuration to apply changes.
- Model streaming responses are empty, and marketing content cannot be generated. Cause:
enable_ocris not enabled, so image-format quotation sheet data cannot be extracted; or recalled knowledge base content is empty; or context length settings are insufficient to trigger streaming output.
How to confirm successful configuration
- Upload a structured table containing steel grade, price per ton, and inventory, plus an image-format quotation sheet. Verify that the knowledge base parsing result correctly extracts the target fields, and field matching accuracy meets business preset threshold requirements.
- Initiate a marketing content generation test based on steel trade inventory and quotations. Verify that the model output covers the configured recall range and context length requirements.
- View the FastGPT system logs. Confirm that model call parameters match the configured items, and there are no 400 errors, timeouts, or other error messages.
- Test the streaming response function. Confirm that generated content is returned in expected segments, with no empty responses or interruptions.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.