What the Data for This Category Looks Like
Data sources for comprehensive service intelligent due diligence reports include public regulatory disclosure documents, enterprise due diligence working papers provided by partner institutions, industry research reports, and compliance review records. The update rhythm adjusts with project progress. Working paper data is updated on demand during a single project cycle. Public information is synced at release nodes. Documents are presented as structured paragraphs plus tables. They include fields such as the subject's unified social credit code, revenue amount (unit: ten thousand yuan), number of compliance penalties, and risk rating. Some documents contain OCR text from multi-page attachment scans.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Decentralized data sources result in mixed text formats, including structured tables and unstructured paragraphs. The recall stage requires distinguishing the weights of field matching and semantic matching. Project-based on-demand updates require support for incremental indexing to avoid repeated full knowledge base construction. Specific fields such as unified social credit code and revenue amount require precise matching, and cannot rely solely on semantic recall. OCR text from attachments has recognition errors. Add a noise filtering step to prevent invalid information from interfering with retrieval results.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the long text paragraph structure of comprehensive service due diligence reports, avoids too short segments breaking semantic coherence, or too long segments causing context loss |
recallTopK | Top 10–15 entries | Covers multi-dimensional compliance, financial, and subject information in due diligence reports, avoids missing key retrieval content |
similarityThreshold | 0.72–0.85 | Distinguishes the boundary between precise field matching and semantic recall, filters low-match irrelevant industry general text |
embeddingModel | Dedicated embedding model that supports long text | Adapts to the characteristic that due diligence reports include OCR text from multi-page attachments, fully captures semantic information of long texts |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Adapts to the parsing time of single due diligence reports with multiple attachments, avoids parsing tasks being interrupted midway |
enableIncrementalIndex | Enabled | Matches the rhythm of project-based on-demand data updates, reduces computing resource consumption of full index rebuilding |
The parameter values provided on this page are all conventional recommendations used to determine the starting point for configuration. Actual values are affected by material form, 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: A new knowledge base gets stuck on the "building index" step with no progress for an extended period. Cause: The
PARSE_FILE_TIMEOUT_SECONDSparameter was not adjusted. The parsing time of OCR attachments for a single due diligence report exceeds the default timeout, causing the task to interrupt. - Phenomenon: Retrieval results do not match the content set in the knowledge base, and irrelevant industry general content is returned. Cause: The
similarityThresholdis set too low, recalling non-target text with insufficient matching, or field-level matching configuration is not enabled. - Phenomenon: A new custom embedding model can complete indexing, but a 500 error is returned during search testing. Cause: The input length limit of the custom model does not match the
chunkSizeconfiguration, and the embedding request during retrieval exceeds the model's supported range.
How to Verify Proper Configuration
- Upload a single complete due diligence report, check whether the parsing progress completes within the time set by
PARSE_FILE_TIMEOUT_SECONDS. - Enter precise field queries, such as "the unified social credit code of a specific subject", verify whether the returned results include accurate information for the corresponding field.
- Trigger an incremental update operation, confirm that only newly added due diligence report data is synchronized to the knowledge base, and no full index rebuilding is triggered.
- Call the retrieval interface, check whether the similarity score of the returned results falls within the preset
similarityThresholdrange.
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.