What the data for this use case looks like
Intelligent due diligence report data for investment platforms mainly comes from public listed company financial reports, industry regulatory disclosure documents, project due diligence working papers provided by partner institutions, and third-party investment research databases. Update rhythms vary by source. Financial reports are updated quarterly and annually. Regulatory documents are pushed in real time. Project working papers are updated dynamically as projects progress. The structure of a single due diligence report usually includes modules such as basic subject information, core financial indicators, risk control and compliance clauses, related transaction details, and investment research rating conclusions. Fields include total assets, operating revenue, and related transaction amount. Corresponding units are yuan, yuan, and ten thousand yuan respectively.
Constraints imposed on multi-turn conversation and prompt engineering
Dispersed data sources and format differences for investment platform due diligence reports require multi-turn conversations to support cross-source field alignment. Prompts must clearly specify field mapping rules for different data sources to avoid data confusion. Differences in update frequencies across sources require conversation context to dynamically filter expired data. Only valid information from the last 30 days is retained. The relatively long length of individual due diligence reports requires multi-turn conversations to limit the proportion of valid content in the context window. This avoids redundant information interfering with core queries. Differences in field units require prompts to mandate that conversation outputs use unified units for calculation results. This meets the standardized needs of investment analysis.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Adapts to the core content length of individual due diligence reports, combined with multi-turn conversation history context, to avoid exceeding model window limits |
recallTopK | Top 3–5 results | Core analysis fields of due diligence reports are concentrated in the top 3-5 retrieval results. Excessive recall introduces redundant information |
promptTemplate | Fixed requirement to output financial indicators in ten thousand yuan units, only retain data sources from the last 30 days, align field mapping rules for different sources | Meets the standardized analysis needs of investment platforms, avoids unit confusion and expired data |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Individual due diligence reports have relatively long length, requiring sufficient time for text splitting and field extraction |
maxConversationRounds | 5–8 rounds | Multi-turn follow-up questions for due diligence analysis usually do not exceed 8 rounds. Excessive rounds lead to redundant context |
similarityThreshold | 0.75–0.85 | Filters low-relevant due diligence document fragments, retains core related content |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing.
Three Common Misconfigurations
- Symptom: After configuring a judgment in advanced orchestration for version 4.6.9 and above, AI conversations fail to retrieve the initial user question. Cause: The
userQueryfield is not bound in the judgment's output variables, causing subsequent AI nodes to fail to read the original query content. - Symptom: The user identification field displayed in conversation logs is empty. Cause: No conversation context user information synchronization rule is configured, and the platform user's unique identifier is not passed to the conversation node.
- Symptom: The due diligence data returned by multi-turn conversations uses inconsistent units. Cause: The prompt template does not clearly specify the requirement for unified unit output, causing the model to randomly use different units for generated results.
How to Verify Configurations Are Correct
- Upload a test due diligence report, initiate an initial query involving financial indicators, and confirm that the conversation context includes the query content.
- Initiate multi-round follow-up questions, and confirm that each reply's context filters out data sources older than 30 days.
- Adjust the similarity threshold, and confirm that the number of retrieved document fragments meets expectations, with no excessive low-relevant content.
- After configuring the prompt template, initiate a query involving unit conversion needs, and confirm that replies uniformly use ten thousand yuan as the output unit.
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.