What This Category of Data Looks Like
Multi-financial financial report data mainly comes from regulatory disclosure platforms and official company announcement channels. Update cadence is divided into regular and temporary content: quarterly reports are disclosed within 45 days after the end of the quarter, annual reports are completed by April 30 each year, and temporary announcements are released in real time as business changes occur. Most documents are in PDF format, including modules such as consolidated financial statements, financial notes, and risk management descriptions. Fields include net trust assets, net finance lease receivables, net fee and commission income, etc. Units are mostly ten thousand yuan or hundred million yuan.
Constraints on Multi-Turn Dialogue and Prompt Engineering
Scattered sources of multi-financial financial report data require multi-turn dialogue to continuously track the latest content from regulatory disclosure platforms and company official websites, to avoid using expired data. Financial report documents have long lengths, with large volumes of content in financial notes and regulatory descriptions. The context window of multi-turn dialogue needs to support long text segment retrieval to avoid truncating the definition descriptions of key fields. Different types of financial reports have distinct update cadences; temporary announcements are released at any time, so real-time verification logic needs to be configured in the interaction flow to ensure each multi-turn interaction uses the latest disclosed content. Industry-specific field definitions have clear standards, so prompts need to clarify field matching rules to avoid model confusion of similar indicators.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 80000-120000 characters | Single multi-financial annual report notes can reach tens of thousands of characters; long context needs to cover financial report segments and user follow-up questions in multi-turn dialogue |
recallTopK | Top 8-12 entries | Multi-financial financial report fields are numerous and scattered; sufficient relevant segments need to be retrieved to cover different dimensions of user multi-turn follow-up questions |
similarityThreshold | 0.75-0.85 | Industry-specific fields have high semantic similarity; low-relevance general financial content needs to be filtered to avoid interfering with interaction results |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Single annual report PDF contains numerous tables and notes; a longer timeout prevents mid-parsing interruptions |
splitChunkSize | 2000-3000 characters | Financial report note paragraphs have moderate length; segmentation facilitates precise retrieval and context matching |
enableSourceInfo | Enabled | Reference knowledge base information needs to be returned in the dialogue interface to meet traceability requirements for multi-turn interactions |
The parameter values provided on this page are general recommendations used as a starting point for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three Common Misconfigurations
- Phenomenon: Full outputs of multiple rounds of AI dialogue are displayed simultaneously after the workflow runs. Cause: Node rules for retaining only final interaction results are not configured, and filtering logic for context truncation is not enabled.
- Phenomenon: A large number of format symbols such as # and * are included in returned content when connecting to external dialogue channels. Cause: The original Markdown format retention switch during knowledge base parsing is not turned off, and format cleaning rules are not configured.
- Phenomenon: Image links in the knowledge base cannot be correctly displayed in dialogue in local deployment scenarios. Cause: The
imageUrlWhitelistparameter is not configured to allow custom URL prefixes, so the model cannot recognize image links not in the whitelist.
How to Verify Proper Configuration
- Upload a single test financial report document, initiate multi-turn follow-up questions, check if the context of each interaction includes the latest financial report segments, and confirm that the
maxContextconfiguration covers the required text length. - Call the dialogue interface, check if the returned results include relevant fields of reference sources, and confirm that the
enableSourceInfoparameter is correctly enabled. - Trigger the temporary announcement update of the knowledge base, initiate relevant follow-up questions, check if the interaction results include the latest disclosed content, and confirm that the knowledge base refresh logic is operating normally.
- Upload a test document containing image links, initiate a dialogue, check if the returned content correctly identifies and displays the image links, and confirm that the
imageUrlWhitelistconfiguration meets the deployment environment 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.