What Data in This Category Looks Like
Investment research data in the multi-finance field falls into two categories: structured and unstructured. Structured data comes from exchange public disclosures and industry databases. It includes fields such as product code, net value, annualized yield, and term. Units include percentage, hundred million yuan, and annualized yield. Unstructured data includes brokerage research reports, industry analysis articles, and private institution quarterly reports. Update frequencies vary by source: research reports are updated weekly or on trading days, while institutional quarterly reports are updated quarterly. Single document lengths vary widely, from hundreds of words of market commentary to tens of thousands of words of in-depth industry research.
What Constraints Do These Characteristics Impose on Citation Source and Traceability
Structured data has high requirements for field accuracy. Traceability must bind specific data release time and release entity to avoid confusing different cycle statistical calibers for the same field. Unstructured documents have wide length ranges. They must support filtering traceability paths by document type and publishing institution. When multiple sources of data on the same topic coexist, traceability hierarchy must be clearly defined to distinguish original data from secondary processed content. Some private institution quarterly reports are semi-public data. Additional permission verification rules must be configured to ensure that data acquisition compliance can be verified during traceability. Traceability for long documents must support precise positioning to the paragraph level. Only document-level traceability cannot meet the precise citation needs of investment research scenarios.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall_top_k | 3–7 entries | Matches the screening needs of multi-source research reports in multi-finance, avoids excessive irrelevant recall results that cause traceability confusion |
source_trace_level | paragraph | Adapts to precise citation of long-form research reports, supports positioning to specific analysis paragraphs |
citation_threshold | 0.75–0.85 | Filters low-similarity recall results, ensures traceability content is highly relevant to queries |
file_parse_chunk_size | 800–1200 characters | Balances semantic integrity and traceability accuracy after long document splitting |
DATA_SOURCE_AUTH_WHITELIST | Broker official websites, industry association, exchange public domain names | Limits legal data sources, ensures traceability content complies with regulations |
citation_display_style | full_source | Displays complete source path, including publishing institution, release time and document link, meets compliance requirements of investment research 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 case-by-case analysis, and it is recommended to test against your own samples before finalizing.
Three Common Misconfigurations
- Unnecessary knowledge base citation markers embedded in generated investment research reports cannot be removed via interface operations. This occurs when the
citation_enableparameter is not configured correctly: full-content citation markers are enabled by default, and no citation filtering rules are set. - Confusing data from different cycles of the same field appears in recall results, and specific release time cannot be distinguished during traceability. This occurs because the timestamp field for data traceability is not bound, data is only associated at the document level, and the release cycle of structured data is not refined.
- Uploaded private institution quarterly report documents are not parsed normally by the system, and institution name and release date cannot be displayed during traceability. This occurs because the domain name of the private institution is not added to the
DATA_SOURCE_AUTH_WHITELISTparameter, and the system blocks parsing of documents from non-whitelist sources.
How to Verify Proper Configuration
- Upload a brokerage research report document, trigger knowledge base recall, and check whether returned results display complete source path including publishing institution, release time and document link.
- Enter a query that includes a specific field name, such as "Annualized yield of a private fund product in 2024", and verify that traceability level of recall results is accurate to specific analysis paragraph.
- Upload a document from a non-whitelist source, check whether system blocks parsing and recall of the document, and verify that
DATA_SOURCE_AUTH_WHITELISTconfiguration is effective. - Adjust
recall_top_kparameter to 2, enter a query related to multiple sources, and confirm that number of recall results matches the set value.
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.