What this category’s data looks like
Energy metals data sources include industry association monthly reports, real-time spot and futures market data from exchanges, public announcements from mining enterprises, supply chain logistics ledgers, and more. Update frequencies are categorized as real-time (futures market data), daily (spot quotes), weekly or monthly (industry research reports), and on-demand (enterprise announcements). Document structures include structured quotation tables (covering metal grades, origin, delivery standards), semi-structured research reports with data charts and industrial chain analysis, and unstructured announcement text. Fields and units include tons, yuan per ton, USD per ounce, grade percentage, and more. Some data requires multi-unit conversion.
Constraints for multi-turn dialogue and prompt engineering
The varied update frequencies for energy metals data require the dialogue system to differentiate data source timeliness. This prevents recalling expired spot quotes. There are many structured fields and unit conversion needs. Prompts must clearly define field matching rules and unit verification logic. This stops confusion between different metal parameter specifications. Long-text research reports need segmentation that retains industrial chain upstream and downstream connections. This avoids breaking data logic when splitting text. High-frequency real-time market data updates require the knowledge base to support incremental synchronization. This prevents using outdated data during dialogue.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Adapts to the context retention needs of long energy metals research reports and multi-turn dialogue, covering upstream and downstream industrial chain related information |
recallTopK | Top 6–8 entries | Balances recall coverage for structured quotation data and semi-structured research reports, avoiding missing key market information |
similarityThreshold | 0.75–0.85 | Accurately matches strongly associated fields such as metal grades and origins, filtering low-relevance non-target category data |
chunkSize | 1000–1500 characters | Retains upstream and downstream industrial chain logic when splitting long research reports, avoiding text breaks that affect analysis |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Adapts to the parsing duration of large structured quotation tables and multi-page research reports, preventing parsing interruptions due to timeout |
| `incrementalUpdateInterval | Once daily | Matches the update rhythm of daily spot data updates and weekly industry report updates, ensuring the timeliness of knowledge base data |
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 perform testing with in-house samples before finalizing configuration.
Three Common Mistakes
- Symptom: After local deployment, uploading a structured quotation file results in an empty processed data list. Cause: The
structuredDataParseparameter is not enabled to enable table parsing, or the structured fields of the uploaded file do not comply with preset specifications. - Symptom: AI dialogue does not reference knowledge base content, only returning general Q&A results. Cause: The
recallTopKvalue is too low, or thesimilarityThresholdis set too high, leading to failure to recall valid documents. - Symptom: No conversation history records appear in the MongoDB collection. Cause: The
enableConversationHistoryconfiguration item is not enabled, or the database connection parameters are configured incorrectly.
How to Confirm Proper Configuration
- Upload a mixed test file containing metal quotations and research report snippets, check if the parsed chunked content retains field and unit information.
- Initiate a test dialogue with multiple follow-up questions, such as first querying "LME copper price" then following up with "Yesterday's price change", confirm that the dialogue context is correctly invoked.
- Call the knowledge base dialogue interface, check if the
sourceDocumentsfield is included in the returned results, and that the field content matches the uploaded document information. - View the conversation history collection in MongoDB, confirm that newly initiated conversation records have been written to the corresponding database and collection.
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.