What the data for this category looks like
Thermal coal due diligence data sources include China Coal Industry Association public statistical materials, northern coastal port spot transaction ledgers, railway transportation dispatch records, and real-time quotes from third-party commodity trading platforms. Update frequencies vary across sources: production area mining data is updated weekly, port spot quotes are updated daily, and industry supply and demand analysis reports are updated monthly. Document structures are primarily mixed format, containing structured due diligence tables with indicators such as production area name, batch number, calorific value, total moisture, ash content, and unstructured regional market analysis text. Indicator units follow industry standard conventions. For example, calorific value uses kilocalories per kilogram, and total moisture uses grams per kilogram.
Constraints Imposed by These Characteristics on Model Access and Configuration
Differences in multi-source data update frequencies require configuring staged incremental synchronization tasks. This prevents temporal bias in model outputs caused by mixing data of different time granularities. Binding of structured fields to specific units requires configuring mapping rules for automatic unit calibration. This prevents model calculation errors caused by unit mismatches across data sources. Mixed-format document structures require configuring parsing priority rules. Prioritize extraction of core indicators from structured tables to reduce redundant interference from unstructured text. Data volume increases with update frequency. This requires configuring reasonable sharding processing rules to avoid model access timeouts or parsing failures caused by overly large single batches of data.
How to Set Configurations
| Config Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | text-embedding-3-large | Thermal coal due diligence data includes multi-dimensional structured indicators. This model has higher discrimination for multi-field vectors, adapting to the multi-attribute retrieval needs of the thermal coal category |
chunk_size | 800–1200 characters | Thermal coal due diligence documents include long-text analysis and structured tables. This range balances context completeness and retrieval accuracy |
SYNC_INCREMENTAL_INTERVAL | Daily data synchronized every 12 hours, weekly data synchronized once per week | Matches the multiple update frequencies of thermal coal data, avoiding temporal bias caused by mixing data of different time granularities |
PARSE_TABLE_STRATEGY | Structured extraction priority | Thermal coal due diligence documents take structured tables as the core data carrier. Prioritizing field extraction reduces redundant information interference |
SIMILARITY_THRESHOLD | Calibrated according to business scenarios | Filters low-match irrelevant data, retaining due diligence information strongly related to thermal coal |
max_context_tokens | Set according to the upper limit of the adapted large model | Avoids input exceeding the model's context limit, preventing call failures |
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 on-site analysis, and it is recommended to test on your own samples before finalizing.
Three Common Misconfigurations
- Phenomenon: After replacing
embedding_model, the retrieval matching accuracy of the existing knowledge base drops significantly. Cause: Vector index reconstruction was not performed. Vector spaces generated by old and new models are inconsistent, making it impossible to correctly match the vector features of existing documents. - Phenomenon: When batch importing thermal coal due diligence reports, the interface displays the
PARSE_FAILEDerror status. Cause:chunk_sizewas not set to adapt to long documents. Single text blocks exceed the parser's processing limit. - Phenomenon: When calling a large model to generate due diligence summaries, the returned results do not include core price indicators. Cause: The
RECALL_TOP_Nparameter was not adjusted. Too many non-core field data were recalled, interfering with the model's core information extraction logic.
How to Confirm Configuration is Successful
- Check the
embedding_modelparameter on the model configuration page. Confirm it matches the currently used vector model, and verify that the parameter name matches official documentation. - Upload a single thermal coal due diligence document. Check if the parsing result fully extracts core indicators from the structured table, and confirm that the
PARSE_TABLE_STRATEGYsetting takes effect. - Initiate a retrieval test. Check the number of returned recall entries and matching degrees, and adjust
SIMILARITY_THRESHOLDto meet business needs. - Test long document input. Confirm that the model call does not return context length-related errors, and verify that the
max_context_tokenssetting is reasonable.
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.