What the data for this category looks like
Data sources for industrial metals research reports include spot and futures prices and inventory data from the London Metal Exchange, Shanghai Futures Exchange, and domestic non-ferrous metal industry associations. They also include in-depth industry research reports and monthly supply and demand reports published by brokerages. Update cadence varies: spot price data updates daily, monthly supply and demand reports release monthly, broker research reports update irregularly alongside industry trends, with intensive updates following policy announcements. Single documents typically include core logic, supply and demand balance sheets, and price trend analysis. Fields cover prices, inventory, production, and consumption across different markets. Units include USD/ton, CNY/ton, 10,000 tons, and similar units.
What constraints these characteristics impose on knowledge base retrieval and recall
Data sources are scattered, and update frequencies differ widely. Support for multi-source access and incremental updates is required to avoid resource consumption from full updates. Research reports contain large amounts of structured numerical values and cross-paragraph logic. When chunking, balance context integrity and matching accuracy to avoid truncating core data. Field units and statistical standards are inconsistent. Complete standardized mapping must be completed before recall to ensure data from different sources can be correctly associated. High-timeliness spot data must recall the latest versions first. Using outdated information risks reducing answer accuracy.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | text-embedding-3-small | Industrial metals research reports contain multi-dimensional structured numerical values and industry logic. This model has stronger adaptability to semantic encoding of numerical text, improving matching accuracy. |
chunk_size | 800–1200 characters | Core supply and demand data paragraphs in industrial metals research reports usually fall within the 500–1000 character range. This length preserves complete data context and avoids truncating key numerical values. |
chunk_overlap | 150–200 characters | Supply and demand logic in research reports has cross-paragraph associations. The overlap interval preserves cross-paragraph context and avoids logical breaks affecting matching results. |
recall_top_k | Top 8–10 results | Data sources for industrial metals research reports are scattered. A sufficient number of candidate results must be recalled to cover supply and demand data from different channels, avoiding missing key information. |
similarity_threshold | 0.72–0.78 | Numerical indicators in industrial metals research reports have strong correlation. A threshold that is too low introduces irrelevant documents, while a threshold that is too high misses relevant data for segmented categories. Calibrate based on actual testing. |
rerank_top_k | Top 3–5 results | Keep the most relevant research report fragments for answering. Excessive redundant information risks interfering with core conclusion output. |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing values.
Three common misconfigurations
- Phenomenon: When calling the knowledge base interface, recall results mix research reports from non-industrial metal categories such as precious metals and steel. Cause: No similarity threshold is configured for industrial metal-specific fields. The threshold is set too low, causing text from unrelated categories to be incorrectly matched.
- Phenomenon: Question and answer accuracy decreases after multi-round conversations. Normal answering resumes after starting a new session. Cause: Context window length is not restricted. Redundant context accumulated across multi-round conversations interferes with semantic matching logic, causing recall results to deviate from core requirements.
- Phenomenon: The knowledge base embedding model automatically switches from
text-embedding-ada-002totext-embedding-3-small, and the ranking of existing recall results changes. Cause: The embedding model version is not locked. The platform updates the model by default without retaining the original configuration, causing changes to encoding logic.
How to confirm proper configuration
- Upload one industrial metals monthly supply and demand report. Check if segmented text retains complete numerical fields such as inventory and production, with no critical content truncated.
- Input specific industrial metals-related questions. Verify that recall result sources cover official channels such as industry associations and broker research reports, with no irrelevant content mixed in.
- Adjust the similarity threshold. Compare recall result relevance across different thresholds to confirm the threshold adapts to the numerical matching logic of industrial metals research reports.
- Trigger an incremental update task. Check if newly released industrial metals research reports are included in the knowledge base index within the set time frame.
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.