Multi-turn Dialogue and Prompt Engineering for Coking Coal Intelligent Due Diligence Reports

Coking coal-related data primarily comes from production ledgers of mines in major domestic producing areas, spot transaction ledgers of northern

What the data for this category looks like

Coking coal-related data primarily comes from production ledgers of mines in major domestic producing areas, spot transaction ledgers of northern ports, delivery settlement data from futures exchanges, and supply and demand surveys from industry self-regulatory organizations. Three update schedules apply. Spot transaction data updates daily. Delivery settlement data updates immediately upon completion of delivery. Industry survey data updates monthly. Most documents are structured tables, with fields including origin identifiers, delivery grade parameters, spot transaction prices, total port inventory, upstream and downstream operating rate values, cross-regional transportation mileage, and more. Units include yuan/ton, ten thousand tons, kilometers, and others.

What constraints these characteristics impose on multi-turn dialogue and prompt engineering

The multi-dimensional update schedule of coking coal data requires dynamic differentiation of real-time and historical data call priorities during multi-turn conversations, to avoid returning outdated information. The large number of segmented fields requires prompts to explicitly specify the range of fields to retrieve, to prevent the model from generating irrelevant content. Format differences across data sources require multi-turn conversations to guide users to supplement specific dimensional information, such as distinguishing between port spot prices and producer ex-factory prices. Monthly updated industry survey data requires clear prompts in conversations about the latest available data cycle, to prevent users from accessing expired content. At the same time, multi-turn conversations must retain context association, to ensure that users’ continuous follow-up questions receive coherent responses.

How to set the configurations

Configuration ItemRecommended ValueRationale
maxContext8000–12000 charactersMulti-turn dialogue for coking coal intelligent due diligence must retain industry data and user requirements from multiple follow-up queries. This length covers full context
retrieval counttop 6–8 entriesCoking coal data includes multiple segmented fields such as prices, inventory, and transportation. A sufficient number of knowledge base entries must be retrieved to cover requirements
similarity threshold0.72–0.80Coking coal has many professional terms. This range balances retrieval relevance and coverage, to avoid omitting segmented field data
segment length1000–1500 charactersCoking coal data has dense structured fields. Segmentation must retain field association, to avoid splitting that breaks data logic
PARSE_FILE_TIMEOUT_SECONDS300 secondsIndustry survey documents are usually lengthy. Sufficient time must be reserved for structured parsing
REMOVE_THINK_TAGenabledThinking tags generated by the model will interfere with due diligence report formatting. They must be automatically removed

The parameter values provided on this page are standard recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.

Three common mistakes

  • Symptom: After the workflow runs, the AI reply still contains content wrapped in <think> tags. Cause: The REMOVE_THINK_TAG configuration item is not enabled, or the code node is not triggered in the production environment.
  • Symptom: Knowledge base answers are truncated, and complete coking coal due diligence field content cannot be output. Cause: maxContext is set too small, or segment length is unreasonable, leading to context overflow.
  • Symptom: The generated due diligence report does not include the specified coking coal delivery grade data. Cause: Retrieval count is set too low, or the prompt does not explicitly limit the range of segmented fields to retrieve.

How to confirm correct configuration

  • Initiate multi-turn follow-up questions covering coking coal spot prices, port inventory, and cross-regional transportation costs. Verify that the reply fully covers the queried dimensions and contains no irrelevant content.
  • Upload a standard coking coal industry survey document. Check that the parsed text segments retain field association, with no logically broken split results.
  • Trigger a workflow that includes a code node. Verify that the AI-generated content contains no content wrapped in <think> tags, and the output format meets due diligence report requirements.
  • Modify the similarity threshold parameter. Compare the knowledge base entries retrieved in two attempts, to confirm that the relevance of retrieval results meets business judgment standards.

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.