What this category of data looks like
This category of data comes from production capacity reports of coal chemical enterprises, industry compliance documents, upstream and downstream supply and demand materials, and marketing adaptation materials for finance, insurance, and wealth management industries. Update cycles are adjusted with quarterly enterprise marketing activities. Industry policy documents are updated per regulatory requirements. Industry research documents are updated per industry cycles. Most documents are formal, chaptered text. Some are scattered marketing script snippets. Fields include product process parameters, capacity indicators, compliance requirement descriptions, and marketing scenario descriptions for financial adaptation. Units include tons, cubic meters, and annual capacity-related units.
What constraints do these characteristics impose on knowledge base retrieval and recall
Coal chemical marketing data for finance, insurance, and wealth management industries has multiple sources. Retrieval requires unified industry terminology and parameter standards to avoid conflicts between process descriptions and financial adaptation scenarios from different sources. A high share of long documents requires reasonable segmentation rules. This prevents excessive single-segment content from causing context overflow. Scattered marketing script snippets require retrieval to match combinations of financial scenario keywords and coal chemical product parameters. This improves recall accuracy. The timeliness of policy documents requires setting appropriate recall thresholds. This prioritizes recently updated compliance and policy-related content. The special nature of unit fields requires matching unit information during retrieval. This avoids results where parameters do not match their units.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Coal chemical documents mostly contain continuous process parameters and compliance descriptions. This segmentation length preserves complete context and avoids separating parameters from their descriptions after splitting |
similarityThreshold | 0.72–0.85 | There is a large semantic difference between coal chemical product parameters and policy texts. This threshold filters low-relevance recall results and improves accuracy |
recallTopK | Top 10–15 results | Coal chemical marketing content requires multi-dimensional matching of product parameters, policies, and scripts. This number of recalled results covers most business scenarios |
rerankTopK | Top 3–5 results | Coal chemical marketing scenarios are relatively single. Retaining the most relevant results after reranking meets output requirements |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Single files of coal chemical industry research documents are large, with long parsing time. This timeout setting ensures complete parsing of long documents |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Single files of coal chemical industry research documents are large. This setting meets large file upload requirements |
The parameter values provided on this page are all common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three common mistakes
- Phenomenon: After knowledge base retrieval completes, AI dialogue output takes more than 10 seconds. Cause: The
recallTopKsetting exceeds 15 entries, and thechunkSizesetting exceeds 1200 characters. This causes the model's input context token count to exceed the threshold and increases inference time. - Phenomenon: Embedding model calls fail, and knowledge base vector generation cannot complete. Cause: The
embedding_api_urlparameter is not configured to point to the interface of the locally deployed m3e-large model. This interrupts the vector generation process. - Phenomenon: Recall results contain content where parameters do not match their units. Cause: Field matching indexes are not enabled, and unit fields are not configured for retrieval. This results in recall entries with incorrect matches.
How to confirm the configuration is correct
- Upload a single coal chemical industry research document. Check if the parsed text segments preserve continuous process parameters and compliance descriptions. Confirm the segmentation rules match the configured values.
- Enter a query related to coal chemical product parameters or financial adaptation scenarios. Check if the number of recalled results and similarity meet the configured thresholds.
- View the knowledge base ingestion logs. Confirm long document parsing and vector generation processes have no timeout or failure prompts.
- Test calls to the embedding model interface. Confirm the vector generation process works normally with no error messages.
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.