What this category of data looks like
Declaration materials and file review data for coal mine production mainly come from mine safety supervision filing documents, production scheduling logs, coal mining and tunneling acceptance files, and gas and roof inspection records. Data update rhythm changes with business nodes: daily inspection records update daily, single working face acceptance files are archived immediately after project completion, and annual or quarterly declaration materials are updated in concentrated batches. Document structure includes fixed fields: mine number, working face ID, declaration date, review status, and hidden danger rectification closed-loop records. Most data units use standard coal industry measurement units such as tons, cubic meters, and hours.
What constraints these characteristics impose on knowledge base retrieval and recall
Coal mine production declaration and review data sources are scattered, including structured logs, semi-structured files, and unstructured inspection records. This requires retrieval to support field alignment logic for multi-format data. Update rhythms vary widely across data types. Daily inspection data requires frequent incremental synchronization, while batch declaration materials require regular full refresh. This imposes constraints on the knowledge base update trigger mechanism. Single file lengths vary greatly. Short inspection records are hundreds of characters, while long acceptance reports are thousands of characters. This requires retrieval chunking strategies to adapt to different lengths of document segments. Hierarchical associations between mines and working faces require retrieval to filter recall results by business dimensions, avoiding invalid cross-mine data from being returned.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Adapts to semantic similarity matching for professional terms in coal mine files. Avoids introducing irrelevant inspection records with too low a threshold, and missing compliant declaration materials with too high a threshold. |
limit | Top 12 entries | Covers multi-working-face data involved in a single declaration material, while controlling context length within model input limits. |
usingReRank | Enabled | Performs secondary ranking on recalled file segments, filtering out hidden danger rectification records with low relevance to the review topic. |
datasetSearchUsingExtensionQuery | Enabled | Expands search terms based on coal mine professional terminology. For example, expands "gas inspection" to "gas concentration monitoring records" to improve recall breadth. |
collectionIds | Specify by target mine dimension | Limits the retrieval scope to the knowledge base collection of the target mine, avoiding invalid cross-mine data interfering with the review process. |
searchMode | Hybrid retrieval | Combines keyword matching and semantic retrieval, adapting to mixed data of structured fields and unstructured text. |
Three Common Mistakes
- Phenomenon: After adding a knowledge base, calling the question answering interface returns an exception, or the interface shows a knowledge base configuration error prompt. Independent testing of knowledge base retrieval returns correct results. Cause: The context splicing rules between the knowledge base and the main model were not configured correctly, leading to overly long file segments exceeding the model's input limit.
- Phenomenon: After switching the vector model, the interface progress bar stalls with no updates, and the old model cannot be restored through original operations. Cause: The vector index cache generated by the old model was not cleared, and the forced index reconstruction process was not triggered.
- Phenomenon: Retrieval results miss some target documents, or contain a large number of file data unrelated to the current review mine. Cause: The
collectionIdsconfiguration was not used to filter the retrieval scope by mine dimension, or thesimilaritythreshold was set too high, missing documents with low similarity but meeting review requirements.
How to Verify Correct Configuration
- Enter the knowledge base retrieval test page, enter a query term containing the target mine number, and confirm that the returned results only include relevant documents of the target mine, and the number of results matches the
limitconfiguration value. - After switching the vector model, check the vector index reconstruction progress. Confirm that it completes within 10 minutes per actual measurement calibration, and that the interface allows switching back to the original model.
- Call the question answering interface with a professional query term, such as "2024 hidden danger rectification status of this working face", and confirm that the returned results include knowledge base recalled file segments, with no invalid cross-mine data.
- After enabling
usingReRank, verify the sorting logic of the retrieval results, confirming that documents with high relevance to the review topic appear at the top.
The values provided here 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.