What data for this category looks like
Intelligent due diligence report data for investment platforms comes from public industry research reports, listed company financial reports, regulatory compliance announcements, offline due diligence working papers, and other sources. Update cadence varies: financial reports are updated quarterly and annually, research reports are released irregularly tied to market trends, and regulatory announcements are updated in real time. Document structure includes both structured fields and unstructured text. Structured fields cover target security codes, industry classifications, financial metrics, and more. Unstructured text includes due diligence process records, risk warnings, valuation analysis, and more. Single document lengths vary widely.
Constraints on vector models and indexing workflow
These characteristics impose the following constraints:
- Mixed index configuration is required, as the data includes structured target fields and unstructured text. Generate and associate vectors separately for structured fields and unstructured text.
- Incremental index triggering logic must be supported, due to uneven update cadence. This avoids reprocessing historical data.
- Chunk sizing must be adjusted flexibly, as single document lengths vary widely. This prevents vector generation errors from overly long chunks, or loss of semantic association from overly short chunks.
- A domain-adapted vector model must be selected, due to specialized financial terminology. This improves semantic matching accuracy.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Matches the paragraph length of standard due diligence reports, avoids overly fragmented chunks that lose semantic association, or overly long chunks that cause vector generation errors |
chunk_overlap | 50–100 characters | Connects context between adjacent chunks, prevents due diligence logic across paragraphs from being split |
embedding_model | General vector model fine-tuned for financial domains | Adapts to specialized financial content such as industry terminology and financial metrics in due diligence reports, improves accuracy of vector semantic matching |
index_type | HNSW index | Balances recall speed and retrieval accuracy for large-scale due diligence data, adapts to the high-frequency query needs of investment platforms |
recall_top_k | Top 10–15 results | Covers multi-dimensional information required for due diligence, avoids excessive irrelevant content interfering with judgment, or too few results missing critical risk points |
incremental_index | Enabled | Adapts to uneven update cadence of financial reports, research reports and other materials, reduces resource consumption from reindexing historical data, improves processing efficiency |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Errors
- Phenomenon: After upgrading the version, uploading CSV-format due diligence working papers throws an error with the prompt
invalid chunk data. Cause: The new version updates the CSV field delimiter verification logic. When thecsv_delimiterconfiguration item is not specified, the default delimiter changes, causing chunking failure. - Phenomenon: The knowledge base indexing task stays in "Indexing" status for a long time with no progress updates. Cause: Incremental indexing is not enabled. Full indexing processing of large-scale historical due diligence data exceeds the system's default timeout threshold, and the task is interrupted.
- Phenomenon: The recall results contain a large number of irrelevant industry research report contents that do not match the due diligence information of the target asset. Cause: No vector filtering rules are configured for the target code field, so the recall scope is not limited to the specified target assets.
How to confirm the configuration is properly set
- Upload a standard due diligence report sample, check that the length of chunked results falls within the preset
chunk_sizerange, with no obviously overly long or short chunks. - Initiate a query for a single target asset, verify that the number of recall results matches the
recall_top_kconfiguration, with no excess or insufficient results. - Import an updated due diligence file, check that the system only indexes the new content and does not reprocess historical files.
- View the vector generation logs, confirm that the model used matches the
embedding_modelspecified in the configuration, with no failed model call records.
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.