What the data for this category looks like
Data for refining and chemical research reports originates from petrochemical industry research institutions, securities firm chemical research teams, publicly disclosed reports of refining and chemical enterprises, and industry exhibition white papers. Update cycles cover fixed schedules and emergency scenarios. Regular quarterly and annual in-depth reports are released at the end of quarters and year-end. Temporary reports covering events such as plant overhauls and oil price fluctuations have irregular update frequencies. Document structures include modules such as core process parameters, market supply and demand data, cost calculations, policy impacts, and future outlooks. The word count of individual documents varies widely. Fields include "device type", "processing conversion rate", "unit energy consumption", "product grade" and other relevant fields. Units mostly use professional metering standards such as tons, kilograms of standard oil per ton of processing volume.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source data with inconsistent update cycles requires the knowledge base to support incremental synchronization and multi-source data mapping configuration, to avoid resource waste caused by full re-import. Complex document structures and dense professional terminology require the retrieval link to support precise segmentation and field-level recall, to avoid splitting critical process logic or confusing professional terms. The presence of professional measurement units requires the vector model to adapt to vector alignment of refining and chemical industry-specific terms, to reduce the probability of low-relevance recall. The wide variation in long document lengths requires segmentation parameters to cover complete business units, to avoid context breaks affecting retrieval accuracy.
How to set configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
CHUNK_SIZE | 800–1200 characters | Each segment of refining and chemical research reports must cover complete process parameters or market analysis units, to avoid splitting critical logic |
RECALL_TOP_K | Top 15–20 results | Refining and chemical research reports have dense professional terminology. Sufficient candidates must first be recalled before reranking to filter results, to avoid missing relevant content |
SIMILARITY_THRESHOLD | 0.72–0.82 | Higher thresholds are required for professional term vector matching, to filter low-relevance general chemical documents |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Sufficient time must be reserved for long document parsing, to avoid timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports uploading of single large annual research reports |
RERANK_TOP_K | Top 5–8 results | Refines recall results to adapt to context length limits for downstream question answering |
The parameter values provided on this page are conventional recommendations used as starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration mistakes
- Symptom: After upgrading to version 4.8.18, the same query fails to recall knowledge base content that was previously matchable. Cause: The vector database index structure was updated during version iteration, and old index data migration or re-import of historical knowledge bases was not performed.
- Symptom: After private deployment using docker compose, a 500 status code error appears when clicking the knowledge base configuration page. Cause: The local access address and port of the vector database were not correctly configured in the default settings, causing the service to fail to connect to the vector database instance.
- Symptom: Knowledge base question answering responses are slow when using a GPU-deployed Milvus vector database. Cause: GPU index acceleration parameters for the vector database were not enabled, or the number of recalled entries exceeds the GPU video memory capacity.
How to confirm configurations are correct
- Upload a typical refining and chemical research report document, check the field integrity and unit consistency of the parsed segments, confirm no truncation or field loss occurs.
- Initiate a query containing refining and chemical professional terminology, verify the number of recall results and similarity scores, adjust parameters to a range that meets business requirements.
- Start the vector database service and initiate a connection test, confirm no connection timeout or permission errors occur.
- Trigger an incremental synchronization task, check whether newly uploaded research reports are included in the retrieval scope.
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.