Multi-turn Dialogue and Prompt Engineering for Minor Metals Investment Research Knowledge Base Construction

Minor metal data sources include domestic spot trading platform quotes, futures exchange market data, monthly survey data from industry associations

What the data for this category looks like

Minor metal data sources include domestic spot trading platform quotes, futures exchange market data, monthly survey data from industry associations, public production capacity announcements from mining enterprises, and operating rate statistics from downstream processing enterprises. Update frequencies vary: spot quotes are updated daily, futures market data is pushed in real time, and industry reports are updated weekly or monthly. Most documents are in structured table format, with fields including product code, origin, daily average price, total inventory, import and export volume, downstream application proportion, and others. Units include yuan/ton, ton, percentage, thousand cubic meters, and more.

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

The difference in update frequencies between spot and futures data requires multi-turn dialogue to distinguish recall priorities between real-time market data and historical reports, to avoid returning outdated data. There are many structured fields and complex units, so prompt engineering needs to clearly specify field extraction rules and unit verification logic, to prevent confusion over measurement standards across different product categories. There are over 20 minor metal sub-categories, so multi-turn dialogue context must bind the target product ID of the current conversation, to avoid mixing data across product categories. Some data has cross-category linkages, so prompt engineering needs to limit the association scope to the specified downstream field, to reduce interference from irrelevant information.

How to set the configurations

Configuration ItemRecommended ApproachBasis for This Approach
maxContextTokens8000–12000 tokenMinor metal data has many fields; a single round of dialogue must carry content from at least 3 structured reports to avoid context truncation
Recall countTop 6–8 entriesThere are many minor metal sub-categories; too many recalled entries will cause context redundancy, while too few will fail to cover the associated data required for investment research
Similarity threshold0.75–0.85Distinguish quote data for different sub-categories within the same category, to avoid mistakenly recalling the price of molybdenum as vanadium market data
Rerank result countTop 3–4 entriesPrioritize returning core data directly related to the current investment research question, reducing interference from non-essential fields
promptTemplateSpliced in the format of "current product + cycle + core indicators"Adapt to the question-asking habits of minor metal multi-dimensional structured data, clarifying the prompt framework
WORKFLOW_MAX_RUN_TIMES1500–2000Address batch call requirements for concurrent investment research conversations, avoiding triggering built-in current limiting rules

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.

Three Common Misconfigurations

  • Phenomenon: Workflow nodes return empty values when concurrent call volume reaches 2-3 calls per second. Cause: The WORKFLOW_MAX_RUN_TIMES parameter is not adjusted to a value suitable for concurrency, triggering built-in current limiting rules.
  • Phenomenon: The result of variable B returned by API calls overlaps with the historical content of variable A. Cause: No variable isolation rules are configured in the workflow, and residual historical session data causes context to not reset.
  • Phenomenon: Preset guide questions are not displayed in the conversation interface after input guidance and thesaurus are enabled. Cause: The "input guidance enable switch" is not turned on, or the thesaurus is not associated with the question and answer module of the current knowledge base.

How to Confirm Configuration is Complete

  • Initiate a single-round investment research query, verify that the returned results include the core fields of the currently specified minor metal product, and that units match the preset rules.
  • Initiate 3 consecutive progressive queries (for example, first ask for the daily average price, then the month-on-month change, and finally the downstream application proportion), verify that the context correctly associates the current product and cycle.
  • Simulate 2-3 concurrent calls per second, check whether the workflow triggers errors or returns empty values, and confirm that the current limiting parameters are suitable for concurrency requirements.
  • View the variable list returned by the API, confirm that each variable only contains independent results of the current conversation, with no overlapping historical data.

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.