What the data for this category looks like
Thermal coal financial report-related data mainly comes from public monthly supply and demand reports released by the China Coal Industry Association, real-time monitoring data from core ports including Qinhuangdao Port, and quarterly and annual regular reports of listed coal enterprises. The update rhythm varies across data sources: industry monthly data is updated monthly, listed company reports are released quarterly and annually, and temporary price adjustment and capacity adjustment announcements are released at any time. The data includes structured numerical fields and semi-structured analysis paragraphs. Fields cover thermal coal calorific value (unit: large calories per kilogram), tax-included car-board price (unit: yuan per ton), port inventory (unit: 10,000 tons), production volume, transportation volume, and more. Data formats from different sources have minor differences.
What constraints do these characteristics impose on the vector models and indexing link
The multi-source nature, varied update rhythms, and specific field units of thermal coal financial report data create multiple constraints for the vector models and indexing link. First, mixed structured and semi-structured data requires embedding models to adapt to both numerical semantics and textual semantics, to avoid vectorization bias caused by unit differences. Second, data with different update frequencies has different timeliness requirements, so recall weights need to be distinguished based on the time dimension. Third, format differences across multiple sources require standardization processing before vectorization, otherwise the semantic matching accuracy of index recall will decline. Fourth, financial report analysis needs to cover multi-dimensional data, so the index must support multi-field joint recall to avoid missing information from a single dimension.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
EMBEDDING_MODEL_TYPE | qwen3-embedding-8b | Adapts to open-source embedding models deployed locally, matches the actual model type used by users |
CHUNK_SIZE | 800–1200 characters | The core information density of single paragraphs in thermal coal financial reports is moderate. This range can retain complete associated semantics of prices and inventory |
RECALL_TOP_N | Top 8–12 entries | Financial report analysis needs to cover multi-dimensional data (production volume, price, inventory). Too many recall entries will increase context pressure, while too few will miss key information |
INDEX_PARTITION_FIELD | update_time | Partitions data by update time, to distinguish recall priorities between monthly industry data and quarterly financial reports |
VECTOR_DB_BATCH_SIZE | 32 entries per batch | Adapts to the video memory usage of locally deployed embedding models, to avoid timeouts for single vectorization tasks |
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, and it is recommended to test on your own samples before finalizing.
Three common errors
- Connection timeout error occurs when deploying locally hosted Qwen3-Embedding-8B via VLLM. The cause is failing to fill in the correct model service address and port on the FastGPT embedding model configuration page, or failing to configure access authentication parameters.
- Unable to query the storage location of chunked documents and vectorization results in the FastGPT knowledge base. The cause is failing to check the file storage directory and vector database mount path specified in the FastGPT configuration file.
- Deviation occurs in knowledge base disk usage statistics, by mistakenly combining three types of storage items for calculation. The cause is failing to distinguish the three independent storage directories of original files, chunked text and embedded vectors, and incorrectly including temporary cache data.
How to confirm the configuration is properly set
- Test embedding model connection: Submit a segment of thermal coal price data text, check if FastGPT returns normal vector embedding results, with no connection timeout or format error prompts.
- Verify index partition configuration: Upload one monthly industry data entry and one quarterly financial report entry, check if independent index partitions are generated in the vector database according to the
update_timefield. - Check disk usage statistics: Enter the server storage directory, view the occupied sizes of original files, chunked text and vector database separately, and match the values displayed in the FastGPT backend.
- Test recall logic: Enter a query related to thermal coal inventory, check if the returned recall results include relevant data from different time dimensions, and the number of entries falls within the
RECALL_TOP_Nconfiguration interval.
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.