What the data for this category looks like
Data for coal mining production internal office and employee services mainly comes from job operating procedures, equipment maintenance logs, safety training manuals, attendance schedules, emergency response plans, and other materials from each post at the mine.
The data update rhythm is split into three categories:
- Equipment maintenance logs and on-site inspection records are updated daily or in real time.
- Safety procedures and training materials are updated quarterly or annually.
- Attendance schedules are updated monthly.
Document structures are divided by department and post, including fields such as document number, applicable post, effective date, risk level, and technical parameters. Technical parameters often use professional units such as millimeters, megapascals, and hours.
What constraints these characteristics impose on knowledge base retrieval and recall
Coal mining documents have many specialized terms and complex unit systems. Retrieval must balance semantic matching and precise keyword matching. Real-time updated maintenance data requires incremental synchronization to the knowledge base to avoid retrieving expired information. The document structure divided by post requires retrieval to filter by permissions. Only available documents for the corresponding post are returned. Long documents create more semantic fragments after chunking. The number and relevance of recall results must be reasonably controlled. In addition, some documents have inconsistent unit formats. Unit formats must be aligned before retrieval, otherwise matching accuracy will decrease.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Coal mining specialized terminology semantic matching requires a high threshold to avoid retrieving irrelevant non-coal mining professional documents |
limit | Top 6–10 results | Coal mining internal documents have abundant professional content. Too many results will exceed the context window, while too few will miss relevant procedures |
searchMode | vector+keyword | Coal mining documents require both semantic matching for specialized terms and precise keyword matching for document numbers and procedure numbers |
usingReRank | Enabled | The professional content of coal mining documents has high similarity. Reranking can improve the accuracy of result relevance ranking |
datasetSearchUsingExtensionQuery | Enabled | Coal mining employees' questions often use colloquial expressions. Extended queries can match colloquial expressions to specialized terms |
collectionIds | Bind corresponding collection IDs by post | Employees of different posts only need to access documents for their own posts to avoid irrelevant information interfering with retrieval results |
Three Common Mistakes
- Phenomenon: Unuploaded content appears in retrieval results, and the LLM generates answers based on non-knowledge base information. Cause: The
similaritythreshold is set too low, retrieving documents that are semantically similar but not part of the target knowledge base, and no secondary filtering is performed viausingReRank. - Phenomenon: Knowledge base query takes too long, the interface shows loading timeout, or the cache configuration does not take effect. Cause:
collectionIdsare not divided by post, performing full retrieval of all types of document data across the mine, and cache parameters are not correctly configured or the cache threshold is set unreasonably. - Phenomenon: Retrieval results include expired safety procedures or invalid equipment parameters. Cause: No filtering is performed based on the document's effective date field, retrieving obsolete old documents.
How to Confirm the Configuration is Correctly Set
- Upload a test maintenance document for a coal mining fully mechanized mining face, initiate retrieval for the corresponding question, and verify that the returned results only include this test document and relevant documents for the same post.
- View the retrieval log to confirm that the number of recalled results does not exceed the value set by
limit, and the reranked results are sorted according to professional relevance. - Modify the
similaritythreshold to 0.6, initiate retrieval, confirm that the number of returned results increases and some non-professional relevant documents appear, verifying that the threshold takes effect. - Switch test accounts for different posts, confirm that only the document collection content bound to the corresponding post can be retrieved.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.