What the data for this category looks like
Data for this category comes primarily from coal mine safety production declaration materials, underground operation case files, safety supervision approval archives, and equipment operation and maintenance records. Updates trigger alongside business workflows. Single declarations are submitted in batches, while daily approval case files are updated on demand. Document lengths vary significantly. Some documents are short job approval sheets spanning a few pages, while others are long specialized production plans spanning dozens of pages. Fields include declaration subject, operation area, equipment number, approval document number, and rectification time limit. Supporting numerical fields use coal-specific units such as tons, meters, and hours.
What constraints these characteristics impose on vector models and indexing
Large differences in document length require chunking strategies that adapt to different splitting logic for short-text approval sheets and long-text plans. This avoids redundant vectors for short texts or semantic breaks in long texts. Multiple fields contain specialized units, so vector models must support vectorization of mixed text and numerical fields. This prevents loss of specialized term semantics. High-concurrency scenarios from batch declaration submissions require indexes to support incremental updates and batch insertion. This avoids timeouts during single insertion operations. Approval document numbers as unique identifiers require indexes to support precise recall by specific fields. This improves retrieval accuracy.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the mixed structure of long-text production plans and short approval sheets in coal mine declaration materials, balancing semantic completeness and indexing efficiency |
| Overlap Length | 100–150 characters | Retains contextual association between split chunks, preventing semantic breaks in long texts across chunks |
| Vector Model Selection | bge-large-zh-1.5 | Matches coal mine specialized Chinese terminology, with high matching degree to the text characteristics of declaration materials |
| Similarity Threshold | 0.72–0.78 | Filters low-relevance case content while ensuring recall accuracy for similar declaration materials |
defaultToken | 4096 | Adapts to the vectorization upper limit of long-text production plans, preventing truncation of key content |
| Maximum Recall per Request | Top 10 entries | Controls information load for manual review, matching the precise retrieval needs of case review |
Three common errors
- Phenomenon: Retrieval results include unconfigured auxiliary documents. Cause: Auxiliary data is not excluded from the index scope, and auxiliary files are mistakenly included in the vectorization and indexing process.
- Phenomenon: Multiple duplicate vector results are returned for a single declaration material. Cause: The overlap length parameter is not configured correctly, leading to repeated splitting of the same text and generation of duplicate vectors, or document deduplication configuration is not enabled.
- Phenomenon: Uploaded local vectorized documents cannot be recalled normally on the server side. Cause: The vector model version and embedding dimension are not unified. The model used on the server side is incompatible with the model used when splitting local documents, leading to vector matching failures.
How to confirm the configuration is correct
- Use the backend interface of FastGPT v4.17.0 to upload a single short-text approval sheet. Check that the number of vector split chunks matches the expected value, and adjust the chunk length to match the document length.
- Import a production plan document containing coal mine specialized terminology. Check that the semantic matching degree of specialized terms in retrieval results meets review requirements, and adjust the similarity threshold.
- Submit batch declaration materials. Check that the index insertion time and system load meet expected values, and adjust the
defaultTokenparameter to adapt to the maximum length of a single document. - Switch the vector model to a version other than bge-large-zh-1.5. Check that the specialized term matching rate of recall results decreases, confirming that the current model configuration is effective.
The values listed above are common starting points. They should be evaluated 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.