Data Characteristics for This Category
Financial report data for the energy storage category comes primarily from periodic reports of listed energy storage enterprises, industry regulatory disclosure documents, and public corporate announcements. Updates are rolled out in batches based on quarterly reports, annual reports, and interim announcements. Most documents are a mix of structured and semi-structured formats, including fields such as revenue breakdowns, installed capacity, unit costs, and R&D investment. Installed capacity units are mostly MW or GWh. Financial fields are mostly denominated in RMB. Some documents include quarterly business detail tables.
Constraints on Vector Models and Indexing
Energy storage financial reports have a high proportion of structured content and include numeric business fields. This requires vector models to support encoding of multiple feature types, avoiding adaptation bias from models optimized only for plain text. Documents include separate business detail tables. When splitting indexes, context associations of tables must be preserved to prevent disruption of business logic. Units are standardized to fixed formats such as MW and GWh. Indexes must link fields to their corresponding units to avoid retrieval confusion. Interim announcements have irregular update cycles. Incremental indexing mechanisms must adapt to non-standard update rhythms. Many long-text business description paragraphs exist. Segmentation must follow business logic boundaries, with no arbitrary truncation.
Configuration Settings
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Energy storage financial reports include long business descriptions and detail tables. This range preserves business logic blocks and avoids truncating critical information |
vector_model | bge-large-zh-v1.5 | This model has strong adaptation to structured text and numeric associated semantics in Chinese financial reports, and supports multi-feature encoding |
index_batch_size | 32–64 | Single energy storage financial report documents have large data volume. This batch size balances indexing efficiency and memory usage |
enable_table_parse | Enabled | Energy storage financial reports include a large number of business detail tables. Enabling this option preserves table structure and field associations |
similarity_threshold | 0.72–0.80 | Semantic similarity of financial report text has high differentiation. This range filters low-relevance retrieval results |
incremental_update_strategy | Triggered by document update time | Energy storage financial reports have irregular update cycles. This strategy adapts to non-standard updates from interim announcements |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- Issue: The exported knowledge base dataset.csv only includes the index field and no content field. Cause: The original text extraction configuration after document parsing is not enabled. Only index metadata is synchronized, and complete business text is not synchronized.
- Issue: When using the bge-large model to import energy storage financial report text, failure to adjust segmentation parameters leads to broken logic in retrieval recall results. Cause:
chunk_sizeis not adjusted for long business paragraphs in financial reports. Default segmentation truncates cross-paragraph business association information. - Issue: Only one GPU is utilized in a multi-GPU hardware environment, and indexing tasks do not use remaining computing power. Cause: The
CUDA_VISIBLE_DEVICESparameter is not configured to specify available GPUs. The system defaults to recognizing only the first GPU device.
How to Confirm Proper Configuration
- Upload a single energy storage financial report document, review the parsed segmentation results, and confirm that segments do not truncate content with business logic associations.
- Enter retrieval terms related to energy storage financial reports, verify the field and unit matching of recall results, and confirm that indexes link fields to their corresponding units.
- Review indexing task logs, and confirm that incremental update tasks can automatically trigger based on the release time of interim announcements.
- Check the GPU mounting configuration of Docker containers, and confirm that multiple GPUs can be properly recognized by the vector model service.
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.