What the data for this category looks like
Data sources include public reports from the China Coal Industry Association, monthly production capacity statistics from industrial and information departments of major producing areas, weekly inventory reports from core ports such as Qinhuangdao Port, and daily position data from futures delivery warehouses. Update frequencies vary by dimension: spot closing prices are updated every trading day, port inventory is updated every 3 working days, and industrial capacity and policy reports are updated monthly. Document structures mostly combine structured tables and industry analysis text, with some exported in PDF or structured Excel format. Core fields include production origin, total moisture, ash content, sulfur content, caking index, colloidal layer thickness, and calorific value. Units include percentage, millimeters, megajoules per kilogram, yuan per ton, ten thousand tons, and other professional measurement standards.
What constraints these characteristics impose on model access and configuration
Coking coal data contains numerous professional terms and standardized measurement fields. Embedded models must have industry semantic recognition capabilities to avoid semantic misinterpretation of professional fields by general models. Update frequencies differ across multiple data sources. Differentiated synchronization cycles must be configured for each data source to ensure data timeliness for due diligence reports. Structured fields account for a large share of the data. Precise segmentation and field extraction rules must be configured to prevent semantic fragmentation caused by splitting professional information units. Due diligence reports need to cover multi-dimensional data. The number and scope of recall results must be limited to avoid redundant information interfering with final output.
How to Configure Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
vectorModel | text-embedding-3-large or bge-large-zh-v1.5 | Coking coal has many professional terms. Large-scale embedding models can better capture industry semantic associations and improve recall accuracy |
chunkSize | 800–1200 characters | Coking coal data contains professional fields and analysis text. This length preserves the integrity of individual professional information units and avoids semantic splitting |
similarityThreshold | 0.75–0.85 | Professional due diligence scenarios require filtering low-relevant industry data. This threshold balances recall coverage and information purity |
recallTopK | Top 6–8 entries | Due diligence reports need to cover core dimensions such as production areas, prices, inventories, and policies. This quantity avoids redundant information occupying context space |
syncInterval | 1 hour (real-time data sources) or 24 hours (batch report data sources) | Matches the update frequency of different data sources to ensure the timeliness of due diligence report data |
parseFieldWhitelist | ["产地", "灰分Ad", "硫分St,d", "粘结指数G", "平仓价"] | Only retains core fields required for due diligence reports, reducing interference from irrelevant data on model context |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Configuration Mistakes
- Phenomenon: The due diligence content returned by the model does not match coking coal professional information, and term misinterpretations occur. Cause: No industry-adapted embedding model is configured. General embedding models cannot accurately recognize the semantics of professional fields.
- Phenomenon: FastGPT fails to connect to the configured large language model, and the interface displays a
504 Gateway Timeouterror. Cause: The model's API gateway address is not configured correctly, or the corresponding port is not opened in the platform firewall, causing requests to fail to route normally. - Phenomenon: When calling the knowledge base conversation interface via HTTP requests, the returned results contain irrelevant chat content. Cause: The model's default chat switch is not turned off, causing the model to generate unintended content outside the incoming knowledge base data.
How to Confirm Successful Configuration
- Input coking coal professional terms to test the embedding model's output, confirm that semantic similarity meets expectations.
- View the execution logs of data synchronization tasks, confirm that different data sources complete updates according to the preset synchronization cycle.
- Initiate a knowledge base conversation request, check whether the number and fields of recalled documents conform to the configured rules.
- Check the model call logs, confirm that each request uses the configured large language model, with no connection errors or parameter abnormalities.
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.