What the data for this use case looks like
Joint-stock bank research report data has two primary sources. First, internal industry analysis reports produced by the bank’s research and development department. Second, public listed company tracking reports from cooperating third-party institutions. Internal reports sync daily on workdays. Third-party reports sync per the publishing party’s official update cycle. Documents use standardized modules. These modules include unified header and footer, title, publishing entity, publish date, covered industries, core conclusions, and data calculation tables. Document fields include: unique report identifier, full publishing organization name, covered industry classification, target enterprise name, investment rating, and target price. Target price uses Renminbi yuan as its unit. Publish dates follow the ISO 8601 standard format.
Constraints for Workflow Orchestration
Mixed internal and third-party report sources require configuration at the workflow start node. Add data source classification rules to prevent cross-source data confusion. High update frequency requires a scheduled trigger scheduling node. Configure the node to automatically sync the latest data on a workday cycle. Uniform document structure with many fields requires custom field extraction rules. Add the rules at the data parsing node to accurately extract core fields. Core fields include investment rating and target price. Some reports contain sensitive internal calculation data. Add a sensitive information desensitization node to ensure compliant data use. Long individual report length requires a segment recall node. Configure the node to avoid context overflow. Context overflow impacts the accuracy of subsequent AI processing.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_CHUNK_SIZE | 800–1200 characters | The body paragraphs of joint-stock bank research reports are moderately long. Segmentation preserves complete logical units and avoids context breaks |
RECALL_TOP_K | Top 6–8 results | Research report datasets are large. Too many recalled results increase context load, too few fail to cover core viewpoints |
SIMILARITY_THRESHOLD | 0.72–0.85 | Research report content is highly specialized. A high similarity threshold is needed to filter irrelevant retrieval results |
WORKFLOW_TIMEOUT | 120 seconds | Research report parsing and recall processes handle multiple text segments. The timeout period must cover the full processing cycle |
SYSTEM_PROMPT | Joint-stock bank investment research assistant, only answer questions based on provided research report content, strictly use rating and target price data mentioned in the documents | Aligns with the usage scenarios of joint-stock bank investment research personnel, avoids generating irrelevant content |
FILE_UPLOAD_MAX_SIZE | 50 MB | The size of individual joint-stock bank research reports (including attached tables) typically does not exceed this threshold, covering most scenarios |
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 targeted analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Phenomenon: After the workflow calls the AI node, the new context field is mistakenly returned as the final output result. Cause: The two output fields of the conversation node are not clearly distinguished by purpose, and the AI reply content field is not connected to the downstream return link.
- Phenomenon: A timeout error with status code 408 occurs when the workflow processes a single research report. Cause: The
WORKFLOW_TIMEOUTparameter is not set appropriately, and the timeout period is shorter than the actual time required for report parsing and recall. - Phenomenon: Non-target industry research reports are mixed in retrieval results. Cause: No data source filtering rules are configured, and valid data is not filtered by the covered industry field.
How to Verify Proper Configuration
- Manually trigger the workflow, upload a single joint-stock bank research report, and verify that the parsed fields include preset content such as investment rating and target price.
- Call the workflow API interface, pass a simulated investment research question, and verify that the returned results are generated only based on the uploaded research report content and do not include external information.
- Configure a scheduled trigger task, wait for one scheduling cycle, and verify that the workflow automatically syncs the latest research report data.
- Send multiple concurrent requests, and verify that the workflow processes requests stably according to configured current limiting parameters without errors.
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.