Multi-turn Dialogue and Prompt Engineering for Coking Coal Investment Research Knowledge Base Construction

Coking coal data primarily comes from public market data of the Dalian Commodity Exchange, industry reports from the China Coal Industry Association

What this category of data looks like

Coking coal data primarily comes from public market data of the Dalian Commodity Exchange, industry reports from the China Coal Industry Association, public operational data from major producing area mining enterprises, and coastal port customs clearance records. Market data is updated daily, with some real-time quotes refreshed every 15 minutes. Industry reports are updated weekly, monthly or quarterly. Corporate financial reports are released quarterly and annually.

Document structure falls into two categories: structured tables and semi-structured reports. Structured market documents include fields such as delivery location, delivery grade, settlement price, and open interest, with units of yuan/ton and ten thousand tons. Semi-structured reports include statistical items such as production capacity, import volume, and transportation costs. Field naming follows industry general standards uniformly.

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

The multi-dimensional segmentation and real-time nature of coking coal data create multiple constraints for multi-turn dialogue and prompt configuration. First, real-time market data has strong timeliness. Prompts must explicitly require recalling documents from the past 24 hours to avoid citing outdated data. Second, investment research dialogues often involve sequential progressive questions. For example, first confirming the main coking coal price at a certain delivery location, then asking about month-over-month changes. Historical dialogue key parameters such as delivery location and grade must be retained, so users do not need to repeat them. Third, coking coal data has high field segmentation granularity. Prompts must enforce matching precise fields such as delivery grade and delivery location. Otherwise, irrelevant data from other coal categories may be included. Finally, single industry reports have long lengths. The total length of context recall must be limited to avoid exceeding the model’s token limit.

How to Configure the Settings

Configuration ItemRecommended SettingRationale
maxContextLast 10 turns of dialogueCoking coal investment research dialogues often involve continuous market and policy tracking. Excessive history dilutes the contextual accuracy of current questions
RECALL_TOP_NTop 8 entriesCoking coal data covers multiple dimensions including market, inventory, and policy. Excessive recall results in an overly long prompt
SIMILARITY_THRESHOLD0.75–0.82Coking coal data has highly segmented fields (delivery location, grade). A threshold that is too low will include irrelevant coal category data
PARSE_FILE_TIMEOUT_SECONDS120 secondsQuarterly coking coal industry reports have long document lengths. Sufficient time must be reserved for parsing
UPLOAD_FILE_MAX_SIZE500 MBHistorical coking coal market database files have large file sizes
WORKFLOW_AI_CHAT_TIMEOUT600 secondsMulti-turn investment research dialogues may involve integrated calculations across multiple documents. The timeout period must be extended

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

Three Common Misconfigurations

  • Symptom: The workflow uses the AI Chat Component and triggers an uncaught exception. The log shows 504 Gateway Timeout. Cause: The PARSE_FILE_TIMEOUT_SECONDS value is not adjusted to suit long coking coal documents, resulting in document parsing timeout.
  • Symptom: Calling the API via Python to initiate a dialogue, and the generated reply does not reference knowledge base documents. Cause: The kbIds parameter is not correctly included in the request body, or the SIMILARITY_THRESHOLD is set too high, resulting in no matching recall results.
  • Symptom: The dialogue page or knowledge base page crashes. The front-end console shows Uncaught TypeError. Cause: The 4.8.20 version with front-end rendering vulnerabilities is used. Upgrade to a stable version.

How to Verify Proper Configuration

  • Initiate a single-turn dialogue test. Ask for the coking coal price at a specified delivery location and grade. Verify that the reply includes precise data for the corresponding fields.
  • Initiate 3 consecutive progressive questions. For example, first ask about main coking coal inventory, then ask about month-over-month changes, and finally ask about influencing factors. Verify that context is correctly retained.
  • Upload a single coking coal industry report. Check that the parsed document fields fully match the original document’s structure and units.
  • Trigger the workflow AI chat component. Check the log for any timeout or parameter error prompts.

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.