What the data for this category looks like
Core data sources for thermal coal include the National Coal Trading Center, coastal port spot trading platforms, and futures exchanges. Spot quotation data is updated daily, monthly supply and demand reports are released each month, and long-term contract transaction data is updated quarterly. Most data documents are in structured table format. Spot documents include fields such as release date, port, and calorific value. Supply and demand documents divide statistical items by region. Long-term contract documents include transaction quantity and execution cycle. Core field units include megajoules per kilogram, mass fraction, and others. The length of a single document varies with data dimensions, and it adapts to the needs of creating marketing content for different customers such as power generation enterprises and traders.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
Because thermal coal data has different update frequencies, multi-turn dialogue must dynamically adjust the recall time range of the knowledge base to avoid using expired long-term contract data when answering spot-related marketing content. Professional field units have strict requirements, so prompts must explicitly specify a unified output unit to prevent the model from performing automatic conversions that cause professional ambiguity in marketing content. Data types are divided into three categories: spot, supply and demand, and long-term contract. Multi-turn dialogue must first confirm the customer scenario the user belongs to, then directionally recall knowledge base content of the corresponding type to improve the targeting of marketing content. The existence of multi-source data requires prompts to clearly specify the preferred data source, to avoid conflicts between data from different sources affecting the accuracy of marketing content.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Thermal coal data documents have a long single-document length, so multi-turn dialogue needs to retain sufficient historical context to avoid information loss |
RECALL_SCORE_THRESHOLD | 0.75–0.85 | There are many professional terms for thermal coal, so low-match irrelevant data needs to be filtered while retaining sufficient professional content |
RECALL_TOP_N | Top 6–8 results | Thermal coal has many data dimensions, so enough relevant field information needs to be recalled to support multi-turn dialogue |
historyMemoryMaxCount | 10–15 turns | In marketing scenarios, users' multi-turn inquiries revolve around specific business needs, and excessive historical content will interfere with current responses |
PROMPT_TEMPLATE | Fixed template, explicitly require using megajoules per kilogram as the calorific value unit, prioritize using the latest data from the National Coal Trading Center, and do not modify the format of knowledge base links | Thermal coal data has strict professional unit requirements, and format errors must be avoided to ensure the accuracy of marketing content |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Batch thermal coal data documents have a large volume, so sufficient time is required to complete parsing |
The parameter values provided on this page are all common recommended starting points for configuration. The actual values are affected by material form, data volume and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing the configuration.
Three common mistakes
- Phenomenon: The knowledge base links output by the model contain extra spaces or incorrect capitalization of letters. Cause: The prompt template does not explicitly require retaining the original link format, and the model's automatic format correction function is not disabled.
- Phenomenon: No response or error status code 408 is returned when calling the file parsing tool. Cause: The
PARSE_FILE_TIMEOUT_SECONDSconfiguration value is too low, and it does not adapt to the parsing duration of batch thermal coal data documents. - Phenomenon: Historical records carry redundant content of designated reply plugins in multi-turn dialogue. Cause: The
historyMemoryFilterrule is not configured to filter plugin execution logs, causing redundant information to interfere with context understanding.
How to confirm the configuration is set correctly
- Initiate a multi-turn dialogue containing thermal coal professional terminology, check whether the calorific value unit in the model output meets the preset requirements, and verify whether the link format is consistent with the original knowledge base.
- Upload a single large-volume thermal coal data document, check whether the parsing status is normal, and confirm that there are no timeout errors.
- Simulate a dialogue process involving plugin calls, check whether redundant content from plugin execution is filtered out of the historical records.
- Adjust the value of
RECALL_SCORE_THRESHOLDto verify whether the matching degree of the recall results meets business needs.
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.