What the data for this category looks like
Thermal coal investment research data mainly comes from monthly supply and demand reports of domestic major producing regions, weekly inventory and price monitoring data from core ports including Qinhuangdao Port, real-time market data from futures exchanges, and statistical analysis documents from industry associations. Update frequencies vary significantly: real-time market data updates at minute-level intervals, port inventory data updates weekly, and monthly industry reports are released in the late part of each month. Document structures include structured tables and unstructured analysis text. Core fields include net as-received calorific value Qnet,ar, total moisture Mt, cleared price, and carboard price. Units include MJ/kg and kcal/kg.
Constraints Imposed on Knowledge Base Retrieval and Recall
Differences in data update frequencies require the retrieval process to distinguish real-time incremental data from historical static reports, to avoid recalling outdated information. A high proportion of structured fields means relying solely on semantic similarity recall cannot accurately match user query demands for specific indicators. Document lengths vary widely: semantic boundaries differ between short market data and long analysis reports. Segment processing must balance completeness and retrieval accuracy. Inconsistent units may lead to chaotic retrieval results, so unit conversion preprocessing must be completed ahead of time.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall Count | Top 15 | Thermal coal investment research needs to cover multi-dimensional data, to avoid missing key indicators or associated information across different reports |
Similarity Threshold | 0.72–0.80 | There are many thermal coal terms and close indicator correlations; a threshold that is too low will introduce irrelevant results, while a threshold that is too high will filter out valid associated information |
Segment Length | 800–1200 characters | Balance semantic completeness of long analysis reports and retrieval accuracy of short market data, avoid truncating key indicators |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Parsing large monthly analysis reports takes a long time, avoid timeout causing file upload failure |
Reranked Return Count | Top 8 | Prioritize displaying the most relevant core data, aligns with the usage habit of investment researchers quickly locating key information |
The parameter values provided on this page are common recommended starting points for configuration setup. Actual values are affected by material form, data volume and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing.
Three Common Mistakes
- Phenomenon: After merging and sorting results from multiple knowledge base calls, the order does not meet investment research requirements, and no arrangement follows the priority of core indicators. Cause: No weighted reranking rules are configured for core thermal coal indicators such as cleared price and calorific value, only basic semantic similarity calculation is used.
- Phenomenon: After uploading XLSX-format thermal coal data, some fields cannot be correctly identified, or units (such as kcal and MJ) are not unified. Cause: Field automatic mapping and unit conversion configuration for table structured parsing is not enabled.
- Phenomenon: When retrieving text and image content related to thermal coal, only text results are recalled without matching corresponding K-line charts or on-site photos. Cause: Cross-modal association configuration for multi-modal retrieval is not enabled, and recall is only performed for text content.
How to Confirm Configuration Is Correct
- Upload a thermal coal monthly report XLSX file that includes core indicators, check if parsed fields are complete and units are unified to the target format.
- Initiate a retrieval query including "thermal coal 5500 kcal/kg cleared price", check if recalled results prioritize the latest port market data and corresponding analysis reports in sorting.
- Configure multiple knowledge base calls (such as real-time market database and historical report database), verify if merged results are sorted and deduplicated according to a unified logic.
- Upload a thermal coal K-line chart, initiate a relevant retrieval, check if corresponding market analysis text can be recalled.
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.