What the data for this category looks like
The data sources for professional services intelligent due diligence reports mainly include public due diligence working papers issued by professional institutions, compliance documents disclosed by regulatory authorities, and research and statistical materials released by industry associations. Update cycles follow project-based workflows. Corresponding data is synchronized immediately after a single due diligence project is completed. General industry due diligence materials are updated quarterly or semi-annually. Most documents follow structured formats, containing fields such as due diligence entity name, unified social credit identifier, financial data, compliance qualification number, and risk warning items. Amount fields default to ten thousand yuan or hundred million yuan as units. Time fields use the YYYY-MM-DD format. Some qualification fields follow fixed-length numbering rules.
Constraints imposed by these characteristics on knowledge base retrieval and recall
The high proportion of structured fields requires retrieval to prioritize exact field matching instead of broad semantic full-text search, to avoid irrelevant content being included in analysis results. The project-based update feature requires the knowledge base to support data isolation by project version, to prevent outdated due diligence reports from being accidentally recalled and harming analysis accuracy. The high proportion of long documents requires text segmentation to preserve field integrity, to avoid breaking semantic connections between financial data or compliance clauses after splitting. Fields with fixed units and formats require unified unit conversion rules during retrieval, to prevent matching failures caused by inconsistent units.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–16000 characters | Single professional services due diligence document has a relatively long length, so it needs to accommodate multiple segments of recalled content and system prompts to avoid context overflow |
Knowledge base recall count | 3–5 entries | Due diligence report data mostly consists of structured exact matches, a small number of recalls can cover the core fields and risk information required for analysis, while too many will increase the pressure of context processing |
Similarity threshold | 0.75–0.85 | Most due diligence report fields are standardized content. A threshold that is too low will introduce irrelevant data, while a threshold that is too high may miss approximately matched compliance or related fields |
Chunk size | 1000–1500 characters | Due diligence reports contain long paragraphs of financial data and compliance clauses. Too long segmentation will lose local semantics, while too short segmentation will damage field integrity |
maxResponseTokens | 2000–3000 characters | Responses related to due diligence reports need to include multi-dimensional field comparisons and risk explanations, so sufficient output space is required to carry complete analysis content |
Rerank result count | 2–3 entries | After rearrangement of structured data, a small number of precise entries can meet the core needs of due diligence analysis, avoiding redundant content interfering with judgment |
The parameter values provided on this page are conventional recommendations for establishing configuration starting points. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Phenomenon: AI-generated replies append fixed brand description text at the end. Cause: The default appended brand statement was not deleted in the system prompt configuration, or the global system append prompt switch was not turned off.
- Phenomenon: After deploying with docker-compose, after a period of time, the knowledge base and workflow configurations become blank, but interface calls can still return normal results. Cause: The configuration directory was not mounted to local persistent storage, and the built-in configuration file was cleared after the container restarted.
- Phenomenon: The matching degree of due diligence report fields returned by retrieval does not meet expectations, and some key financial fields are not recalled. Cause: The similarity threshold was set too high, resulting in only exact matching fields being retrieved, and approximately matched compliance or related data was not covered.
How to Confirm Configuration is Correct
- Upload a standard due diligence report document, trigger the parsing task, and check that the generated field list after parsing matches the preset field structure of the original document.
- Input a retrieval question containing specified financial fields, compliance clauses or entity names, check the relevance of the recalled results, and adjust the corresponding configuration items to meet business judgment standards.
- Trigger a complete question-and-answer process, check that the reply content does not contain additional default appended text, and only includes answers generated based on knowledge base content.
- Restart the deployed container instance, check that the knowledge base and workflow configurations are not lost, and interface calls can normally return expected results.
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.