What the data for this category looks like
Power industry financial report data mainly comes from annual, semi-annual, and quarterly reports regularly disclosed by listed power enterprises, as well as operational statistics data publicly released by provincial power grids. The data update cycle follows a quarterly core rhythm, with annual reports as the full cycle. A single financial report document includes modules such as revenue composition, power generation costs, installed capacity, power generation volume, and grid-connected electricity price. Most fields relate to amount, capacity, and electricity quantity types, with units including ten thousand yuan, ten thousand kilowatts, ten thousand kilowatt-hours, yuan per kilowatt-hour, and similar units.
What constraints these characteristics impose on multi-turn conversation and prompting
The multi-module, long-document structure of power financial reports requires multi-turn conversations to support long context association. This prevents breaks in professional field connections caused by document splitting. Specialized segmented field types require prompts to clearly define analysis dimensions, avoiding irrelevant content generation. Frequently updated quarterly data requires the conversation system to quickly locate the latest disclosed report versions, preventing use of outdated data. Multi-dimensional electricity quantity and amount fields require retaining specific metric anchors from historical conversations during multi-turn follow-up questions, ensuring context consistency for subsequent queries.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–16000 characters | Adapts to the single segment length after splitting power financial reports, retains complete historical context for multi-turn conversations |
RAG_RECALL_TOP_K | Top 6–10 entries | Covers multi-dimensional professional fields in power financial reports including revenue, costs, installed capacity, and power generation volume |
PROMPT_TEMPLATE | Fixed limitation to only answer based on uploaded power financial reports, clearly label the unit of each metric | Avoids generating irrelevant content, unifies unit descriptions for professional fields |
UPLOAD_FILE_MAX_SIZE | 50 MB | Matches the conventional file size range of single power financial report PDF documents |
CONVERSATION_TIMEOUT | 300 seconds | Supports long document parsing and field association calculations during multi-round follow-up questions |
TOKEN_STAT_SWITCH | Enabled | Supports separately counting the number of tokens for conversation input and output |
The parameter values provided on this page are common recommendations used as starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and testing on own samples is recommended before finalizing settings.
Three common mistakes
- Phenomenon: After using Markdown formatting rules in the prompt, the output content is Markdown source code instead of rendered, readable format. Cause: The prompt template does not explicitly require the AI to output rendered text, or the Markdown rendering function is not enabled in the system configuration.
- Phenomenon: Unable to directly reference the financial report metrics mentioned in the previous round during multi-turn conversations. Cause: The
maxContextconfiguration value is insufficient, failing to retain complete historical conversation context, resulting in lost metric anchors. - Phenomenon: The conversation interface does not separately display the number of input and output tokens. Cause: The
TOKEN_STAT_SWITCHconfiguration item is not enabled, so the system does not activate token statistics functionality.
How to confirm the configuration is correct
- Upload a power financial report document of conventional scale, verify that the system can accurately extract and associate professional metrics and their corresponding units within the document.
- Initiate two progressive questions: first specify a financial report module, then follow up with detailed questions based on that module, verify that the system can retain context association from historical conversations.
- Check the conversation-related interface, confirm that there is a display area for separately counting input and output tokens.
- Configure a Markdown-formatted prompt template, initiate a test question, verify that the output content is rendered, readable format.
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.