What Data for This Category Looks Like
Data sources for coking coal due diligence reports include industry analysis reports published by the China Coal Industry Association, spot price systems at major coastal ports, monthly procurement records from domestic key steel mills, and inventory monitoring data from futures delivery warehouses. Update frequencies vary: spot prices are updated daily, industry analysis reports monthly, steel mill procurement records every ten days, and policy documents are updated immediately upon release. Each individual report includes five sections: regional distribution, supply and demand balance analysis, price trends, downstream demand correlation, and policy impact interpretation. The data contains both structured metric entries and unstructured analysis text. Available fields include origin, dry basis ash content, dry basis sulfur content, caking index, colloidal layer thickness, carboard quotation, total inventory, and release date. Price units are yuan/ton, inventory units are tons, and caking index is a dimensionless metric.
Constraints on Vector Models and Indexing
Differences in data update frequencies require indexing strategies that balance scheduled full refreshes and real-time incremental updates. This avoids resource waste or data lag. Mixed structured and unstructured data from multiple sources requires hybrid vector indexing. This covers retrieval needs for both professional metrics and analysis text. Coking coal-specific industrial terms such as caking index and colloidal layer thickness require vector models adapted to domain semantics. This avoids semantic bias from general-purpose models. Long unstructured text in individual reports requires sufficient contextual overlap during chunk processing. This ensures coherence of analysis content. Multi-source data must be mapped to standardized fields. Source identifiers must be retained during indexing to support traceability requirements for due diligence reports.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | bge-large-zh-v1.5 | This model delivers strong performance for Chinese industrial domain terminology, and accurately captures semantic associations specific to coking coal metrics |
chunk_max_length | 800–1200 characters | Matches the length of single-paragraph analysis text in coking coal due diligence reports. Prevents embedding distortion from overly long text, and avoids broken context from overly short text |
chunk_overlap_ratio | 0.15–0.2 | Retains overlapping content between adjacent chunks, ensures semantic integrity of long text, and supports coherent supply-demand analysis in due diligence reports |
retrieval_top_n | Top 8–12 results | Coking coal due diligence reports cover multi-dimensional metrics, requiring sufficient retrieved relevant chunks to cover different analysis scenarios |
retrieval_threshold | 0.72–0.78 | Filters low-similarity irrelevant data, while retaining weakly relevant matches for coking coal professional metrics |
index_refresh_cron | Daily 02:00 and Hourly 00:00 | Balances real-time updated spot data and periodically updated industry reports, optimizes index resource usage and data freshness |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require targeted analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Configuration Errors
- Issue: No noticeable reduction in vector database query latency when repeatedly querying the same coking coal metrics. Cause: Vector retrieval result caching is not enabled, so embedding and retrieval processes run again for each query.
- Issue: Newly uploaded coking coal data cannot be matched during retrieval immediately after adding it to a collection. Cause: The
create_training_orderinterface was not called to trigger index construction. Only data upload was completed, and no vector index was generated. - Issue: Retrieval results for individual coking coal due diligence reports have broken context, making coherent analysis text unreadable in full. Cause:
chunk_max_lengthis set too small, splitting coherent analysis text into multiple unrelated chunks.
How to Confirm Correct Configuration
- Upload a test coking coal spot price document, check the vector index generation status of the knowledge base, confirm that the index status field shows completed.
- Run two identical coking coal metric retrieval queries, compare the two query latencies, confirm that latency drops significantly after the caching mechanism activates.
- Adjust the similarity threshold parameter, verify that retrieval result relevance changes as expected with the threshold.
- Call the batch training order interface to upload bulk coking coal data, check if corresponding task entries are generated in the index queue.
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.