What the data for this category looks like
Data sources for coke research reports include industry reports from the China Coking Industry Association, coke futures market data from the Dalian Commodity Exchange, in-depth research reports from securities firms, and weekly spot trade reports from major domestic ports. Update schedules follow three patterns: daily spot price and inventory data, weekly industry supply and demand tracking, and occasional in-depth analysis reports. Document structure includes fields such as core viewpoint summaries, core supply and demand indicators, downstream industry operating rates, price trend explanations, and policy impact analysis. Units include yuan/ton, ten thousand tons, thousand cubic meters, and others. Some research reports include structured tabular data.
Constraints on vector models and indexing
The multi-frequency updates and mixed structure of coke research reports create multiple constraints for the vector model and indexing workflow. Daily updated spot data requires high-frequency synchronization. The index must support incremental updates to avoid resource consumption from full reconstruction. Structured supply and demand data coexists with unstructured analysis content. Field semantic weights must be distinguished to prevent generic semantic vectors from diluting the matching accuracy of professional data. Long document splitting must preserve data relevance. For example, a single segment of supply and demand data must not be split into different vector blocks, as this would damage logical integrity. Data with different update frequencies must be indexed separately. This enables separate storage of hot and cold data and optimizes retrieval efficiency.
How to set configurations
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the typical length of supply and demand analysis paragraphs in coke research reports, avoiding splitting that breaks data logical relevance |
chunk_overlap_rate | 10–15% | Retains contextual overlap between adjacent segments, ensuring semantic coherence for structured tables and price trend descriptions |
embedding_model | bge-large-zh-1.5 | Adapts to semantic understanding of Chinese professional terminology, with an embedding dimension of 1024, compatible with most mainstream vector databases, and compatible with v4.8.7 version interface configuration |
recall_top_n | 10–15 entries | Covers the professional content scope of coke research reports, avoiding omission of key supply and demand data or analysis viewpoints |
similarity_score_threshold | 0.75–0.85 | Filters low-relevance unstructured content, retaining research report segments with high semantic matching to search terms |
index_refresh_interval | 1 hour | Matches the update rhythm of daily spot data, ensuring the latest market data is included in the retrieval index in a timely manner |
The parameter values provided on this page are general recommendations for starting point configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Search results include unrelated auxiliary data fields, such as port berth information and research trip records. This occurs because separate vectorization rules for structured fields are not configured, and auxiliary data is included in the full-text vectorization process.
- The interface cannot support multi-vector association configuration, and returned results have confusing matching logic. This occurs because multi-vector index configuration is not enabled. Only a single vector model is used to bind all document blocks, and scenarios where one set of data corresponds to multiple sets of vectors are not handled.
- A dimension mismatch error appears after uploading local vector files, prompting that embedded vectors cannot be loaded. This occurs because the embedding dimensions of the vector model used locally and on the server are not aligned. Vector output dimensions vary between different models.
How to confirm correct configuration
- Upload a single standard coke research report, check the number of generated vector blocks, and confirm that segment lengths match the configured parameters.
- Enter search terms related to coke prices or supply and demand, view the associated fields of returned results, and confirm that structured data is correctly recalled.
- Upload updated spot data, check the index update log, and confirm that the incremental update process is triggered normally.
- After replacing the vector model, verify that uploaded vector files can be loaded normally, and confirm that embedding dimensions meet requirements.
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.