What the data for this category looks like
Sources of refractory material research reports include industry allocation reports released by financial institutions, operation briefings published by industry associations, technical research and development reports of listed refractory material enterprises, feature articles from building materials professional journals, and application standard documents released by relevant national departments. The update rhythm varies: industry operation documents are updated quarterly, technical standard documents are updated according to the national standard revision cycle, enterprise R&D documents are released irregularly following project progress, and financial institution allocation reports are updated irregularly based on market trends. Document structure usually includes modules such as category classification, core performance parameters, application scenario matching, cost composition analysis, and upstream and downstream supply and demand data. Core parameter fields include refractoriness, bulk density, and thermal conductivity, with corresponding units of degrees Celsius, grams per cubic centimeter, and watts per meter-kelvin respectively.
What constraints do these characteristics impose on the vector models and indexing link
The professional parameters of refractory material research reports are dense, terms are subdivided, and they are tied to specific application scenarios, which brings multiple constraints to the vector model and indexing link. First, a large number of performance parameters with fixed units require the embedding model to accurately capture professional semantics; general embedding models struggle to distinguish subtle differences between similar parameters. Second, the update rhythm of research reports is uneven, with quarterly operation data and irregular R&D documents coexisting, requiring the index to support incremental updates to avoid resource consumption from full reconstruction. In addition, some documents include performance test charts, which require additional adaptation of the image indexing module to extract structured parameters within the charts, further improving recall accuracy. At the same time, the strong binding relationship between parameters and application scenarios requires the chunking strategy to prioritize splitting by functional modules, rather than splitting by fixed length which breaks semantic associations.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
EMBEDDING_MODEL | bge-large-zh-v1.5 (or vertically fine-tuned embedding models for the domain) | This model delivers stable semantic embedding effects for Chinese industrial and building materials professional terms, and can accurately match the association between refractory material parameters and scenarios |
CHUNK_SIZE | 800–1200 characters | The core parameter modules of refractory material research reports mostly fall within this length range; splitting according to this range avoids breaking the semantic binding between parameters and application scenarios |
RECALL_TOP_K | Top 8–12 results | The number of associated documents for relevant parameters in a single research report is moderate; this value balances recall coverage and context redundancy |
PARSE_TABLE_ENABLE | Enabled | Refractory material research reports contain a large number of performance parameter tables; enabling this setting allows complete extraction of structured data within tables for vectorized indexing |
IMAGE_INDEX_ENABLE | Configured per version | The commercial edition v4.9.0 and above supports image indexing by default; local deployments need to confirm whether the corresponding OCR and image embedding models are loaded |
EMBEDDING_BATCH_SIZE | 32–64 | Balances the processing speed of vector generation and GPU memory usage, and is suitable for deployment environments with single GPU memory of 16GB or higher |
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 conduct tests on your own samples before finalizing the settings.
Three common mistakes
- Issue: When creating a new knowledge base in a local deployment of version v4.9.0, no image indexing model option is available. Cause: The local deployment package for this version does not integrate the OCR and embedding models required for image indexing by default; corresponding plugin packages must be manually loaded.
- Issue: Vector generation fails when
EMBEDDING_API_TYPEis configured as OneAPI, with OneAPI returning connection timeout or model not found errors. Cause: The embedding model endpoint and key for OneAPI are not configured correctly, or the selected embedding model has not been deployed on the OneAPI platform. - Issue: The relevance of refractory material parameters recalled after index generation is low, and the number of results deviates significantly. Cause: The
CHUNK_SIZEis not adjusted according to the module characteristics of refractory material research reports, using fixed short-length splitting which breaks the association between parameters and scenarios, or theRECALL_TOP_Kvalue is unreasonable.
How to confirm the configuration is correct
- Access the knowledge base management interface, view the parsing details of uploaded refractory material research reports, and confirm that table content and image text (if any) have been correctly extracted.
- Test the vector generation API call by inputting a segment of refractory material parameter text, and check whether the returned vector data format meets the preset standards.
- After configuring the embedding model accessed via OneAPI, initiate a small-scale vector generation test to confirm that there are no connection errors or model not found prompts.
- Retrieve preset refractory material professional terms, verify the relevance and number of recalled results, and adjust the corresponding configuration items to values that meet business requirements.
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.