What the Data for This Category Looks Like
Coking coal marketing-related data primarily comes from spot trading ledgers, industry supply and demand weekly reports, customized marketing script libraries, offline roadshow PPTs, and customer communication records. Data update cadence varies significantly by material type: spot price data is updated daily, supply and demand analysis data is updated weekly, and customized marketing content is updated per project milestones. Document formats include two categories: plain-text marketing scripts and quotation documents with structured tables. Core fields include coking coal delivery grade, ash content, sulfur content, origin, tax-included unit price (unit: yuan/ton), and target customer group tags.
What Constraints These Characteristics Bring to Vector Modeling and Indexing
The multi-type characteristics of coking coal marketing data impose multiple constraints on vector modeling and indexing. Structured numerical fields such as ash content, sulfur content, and unit price must be vectorized alongside unstructured marketing text, requiring adapted feature extraction logic for different modalities. Varying update frequencies across different materials require indexes to support flexible switching between incremental and full updates. Daily updated spot price data requires a near-real-time index trigger configuration. Documents with structured tables must support accurate extraction and vectorization of table content, to avoid information loss caused by extracting only plain text. The wide range of raw values for continuous numerical fields requires normalization configuration for vector models to align feature scales, preventing deviations in similarity calculations.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Coking coal marketing documents include long-form marketing scripts and short structured fields. This range balances semantic completeness and complete field extraction, avoiding semantic fragmentation |
vector_normalization | Enabled | Corresponds to a new configuration added in version 4.8.23 and above. Coking coal data includes continuous numerical fields such as ash content and unit price. Unnormalized vectors will have reduced similarity accuracy due to scale differences. This setting adapts to embedding models that do not normalize by default |
recall_top_k | Top 8–12 results | Coking coal marketing content covers multiple dimensions including quotations, scripts, and customer group tags. This value balances recall coverage and context load |
index_update_mode | Hybrid incremental + full update | Spot price data is updated daily, so incremental update reduces resource consumption. Supply and demand weekly report data is fully updated weekly to ensure information synchronization |
parse_table_enable | Enabled | Coking coal marketing documents include structured quotation tables. Enabling this setting fully extracts numerical and field information within tables, improving vectorization completeness |
The parameter values provided on this page are general recommendations used as starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require targeted analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- Phenomenon: The knowledge base returns a matching list of coking coal-related documents, but the large language model prompts that no relevant content was found when generating a reply, and does not use the recalled document information. Cause: Recalled documents are not correctly associated with coking coal-specific fields such as delivery grade and unit price, or the context token setting exceeds the processing capacity of the large language model, resulting in valid content not being fully read.
- Phenomenon: Enabling enhanced index configuration does not result in improved retrieval accuracy. Cause: No dedicated vector matching logic is configured for coking coal's structured numerical fields. Enhanced indexing only applies to plain text content, and does not cover similarity matching for key decision-making information such as ash content and unit price.
- Phenomenon: Knowledge base query response is slow, and token consumption per request exceeds normal ranges. Cause:
max_context_tokenis set too high. All recalled document content is fully sent to the large language model, exceeding normal context processing load, leading to increased processing time.
How to Confirm Correct Configuration
- Select a test document containing a coking coal structured quotation table, submit it to the knowledge base, and check if the segmented content after indexing includes the field information within the table, to confirm that the table parsing configuration is active.
- Submit a new coking coal spot price data set, test the trigger logic of the index update task, and confirm that the switching between incremental and full updates matches the preset configuration.
- Retrieve marketing content containing a specific coking coal grade, verify the number of recall results matches the configuration parameters, to confirm that the recall logic is working correctly.
- Check the activation status of the vector model normalization switch, and verify the optimization effect of the normalization configuration on numerical field matching through changes in similarity scores from test samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.