What the Data for This Category Looks Like
Data related to coking coal due diligence primarily comes from the China Coal Industry Association, Dalian Commodity Exchange, major port customs clearance systems, listed coal enterprise financial reports, and Ministry of Ecology and Environment policy documents. Update cycles are divided into four categories based on data source type:
- Futures market data is updated daily
- Port clearance statistics are updated weekly
- Industry monthly reports are released monthly
- Enterprise financial reports are updated quarterly
Document formats include three types: structured Excel price sheets, PDF industry research reports, and web-based policy announcements. Structured document fields include origin, ash content, sulfur content, caking index, colloidal layer thickness, spot price, with units such as percentage, millimeters, yuan/ton, and others. Unstructured research reports contain sections on capacity analysis, policy interpretation, transportation cost calculation, and more.
What Constraints Do These Characteristics Impose on Knowledge Base Retrieval and Recall?
The multi-source, heterogeneous nature of coking coal data requires the retrieval system to support both structured fields and unstructured text. It must avoid disconnecting professional terms from their corresponding indicators. Data sources with different update frequencies require differentiated recall priorities. Futures market data must recall the latest daily content first. Monthly industry reports must be updated to the knowledge base on a fixed cycle. The uniqueness of professional fields requires precise term matching during retrieval. For example, "caking index" must not be confused with "volatile matter". Confusing these terms will lead to deviations in due diligence data. The chapter structure of long documents requires retaining context association during segmented retrieval. This prevents loss of the binding relationship between policy interpretation and corresponding indicators due to overly short segments. The similarity of cross-category data (such as price fields between thermal coal and coking coal) requires adding category tag filtering during retrieval. This prevents irrelevant data from mixing into results.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
PARSE_SEGMENT_LENGTH | 800–1200 characters | Coking coal due diligence reports contain professional term combinations and long sentences. Excessively long segments will lose context association, while excessively short segments will break the binding relationship between indicators and interpretations |
Recall count | Top 8–12 entries | Coking coal data covers multi-dimensional content including spot, futures, policy, capacity and more. Too few entries will miss critical data sources, while too many will increase context redundancy |
Similarity threshold | 0.75–0.85 | Coking coal professional terms have high recognizability. A threshold that is too low will introduce irrelevant category data such as thermal coal, while a threshold that is too high will miss valid documents for similar indicators |
Rerank result count | Top 3–5 entries | Intelligent due diligence reports need to focus on core indicators. Too many recall results will increase reading and organization costs |
maxContext | 4000–6000 characters | Coking coal data fields are closely linked. Sufficient context must be retained to support joint retrieval across fields (such as ash content and caking index) |
UPLOAD_FILE_MAX_SIZE | 500 MB | Single coking coal industry research reports or batch port data files have large file sizes, requiring adaptation to large file upload requirements |
The parameter values provided on this page are common starting points for configuration setup. Actual values are affected by material format, data volume, and business rules. Specific issues require targeted analysis, and testing against local samples is recommended before finalizing configuration values.
Three Common Misconfigurations
- When uploading knowledge base files via API, Chinese file names display as garbled text. The API returns a 200 status code, but file parsing fails. The cause is failing to specify
Content-Typeasmultipart/form-data; charset=utf-8in the API request header, leading to incorrect transmission of the file name encoding. - When creating a knowledge base, a locally deployed ChatGLM2 model cannot be selected. Retrieval returns empty results. The cause is failing to correctly bind the FastGPT API key in the OneAPI configuration, or failing to configure the model's
base_urlandmodel_nameparameters inconfig.json. - The knowledge base recall results include irrelevant category data such as thermal coal. The number of recalled entries does not match the set value. The cause is a similarity threshold set too low, and failure to add coking coal category tag filtering to documents, leading to recall of cross-category similar data.
How to Confirm Configuration Is Correct
- Upload a single coking coal spot price Excel file, and check that parsed fields include professional fields such as origin, ash content, sulfur content, with no missing fields or garbled text.
- Initiate a retrieval test, input "main coking coal spot price", and verify that the document sources of the recall results are coking coal-related data sources, with no irrelevant category content such as thermal coal.
- View the knowledge base update log, and confirm that daily updated futures data and monthly updated industry reports have been synchronized according to the preset schedule.
- Call the FastGPT V4.8.17 API retrieval interface, and check that the returned result fields include correct coking coal professional indicators, with no abnormal null values.
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.