What the data for this category looks like
Joint-stock bank research report data comes from three main sources. These include industry analysis reports and listed company tracking reports produced by the bank’s research and development department, plus macroeconomic research reports provided by partner institutions. Update schedules cover three types of content: monthly tracking reports, quarterly in-depth reports, and temporary reports for sudden industry events. Most documents use PDF format. Their standard structure includes report title, release date, research subject, core data tables, investment ratings, risk warnings and other fields. Common financial statistical units appear in the data, such as ten thousand yuan, percentage points, and hundred million shares.
Constraints for Multi-turn Dialogue and Prompt Engineering
Long research report documents and included structured tables create specific operational constraints. Multi-turn conversations must retain contextual connections to avoid redundant recalled content. Users frequently request specific data from particular report sections. Prompts must explicitly require linking research topics from prior conversation history. Statistical units vary across different reports. Prompts must explicitly require labeling data sources and their corresponding units. Frequently updated research reports require the dialogue system to prioritize matching the latest released report versions during recall. This prevents use of expired data. Structured table extraction requires enabling corresponding parsing configurations. This ensures accurate location of specific values within tables during multi-turn follow-up questions.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Single research report length ranges from 5000 to 10000 characters. Multi-turn dialogue requires retaining 3 rounds of context, to avoid exceeding model window limits |
recall count | top 6 entries | Joint-stock bank research report knowledge bases have many entries. Too many recalled entries cause context overload. Too few fail to cover complete research dimensions |
similarity threshold | 0.75–0.85 | Research report content is highly specialized. A high matching threshold ensures recalled content is strongly relevant to the query, avoiding irrelevant interference |
PARSE_TABLE_ENABLE | Enabled | Research reports contain many core data tables. Enabling this extracts structured table data, supporting accurate answers to multi-turn follow-up questions |
knowledgeRefreshCycle | 7 days | Monthly tracking reports update every 7 days. Regular refresh ensures recalled content uses the latest released versions |
detail | true | Must be enabled during API calls. This ensures complete reference sources and data details are returned, matching output consistency for online conversations |
The parameter values provided on this page are standard starting points for configuration. Actual values are affected by material format, data volume and business rules. Each specific scenario requires separate analysis. It is recommended to test against your own samples before finalizing settings.
Three Common Misconfigurations
- Phenomenon: API calls return results that differ significantly from online conversations. Even when stream is set to false and detail is set to true, matching results cannot be achieved. Cause: The corresponding research report knowledge base ID is not bound in the API request parameters. This causes the call to not associate preset knowledge base content.
- Phenomenon: The model cannot link research report sections mentioned in previous conversation history during multi-turn dialogue. It returns repeated or irrelevant content. Cause: The
maxContextparameter is not configured to limit context length. This allows redundant historical research report fragments to interfere with current answer logic. - Phenomenon: After configuring prompts, answers do not label data sources and statistical units for research report data. Cause: The prompt does not explicitly require labeling data sources, or the
PARSE_TABLE_ENABLEparameter is not enabled. This fails to extract unit information from research report tables.
How to Verify Correct Configuration
- Initiate a single-turn test conversation. Input a specific data question from a research report. Verify that the returned result includes the corresponding research report reference source.
- Initiate a multi-turn test conversation. Successively ask about different sections of the same research report. Verify that the model can link research topics from previous conversation history.
- Call the API interface to send a request. Verify that the
sourcefield in the returned result includes the research report’s releasing institution and release date information. - View the knowledge base refresh log. Verify that research reports have completed updates according to the preset cycle. Ensure recalled content uses the latest versions.
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.