What the data for this category looks like
Data sources for chemical pharmaceutical intelligent due diligence reports cover pharmaceutical company annual reports, clinical trial submission materials, patent literature, raw material test logs, and CRO outsourced research reports. These are core data sources for financial institutions conducting due diligence on chemical pharmaceutical companies. Update rhythms vary significantly: annual reports update per fiscal year, clinical data updates in phases aligned with trial stages, and raw material logs update in real time per batch.
Document structures include large numbers of structured tables covering clinical trial data, raw material purity parameters, synthesis process parameters, alongside long text for research background and process descriptions. Fields include IC50 values, CAS numbers, batch numbers, purity percentages, with units such as mol/L, %, g and other professional measurement identifiers.
What constraints do these characteristics impose on the document parsing and chunking link?
The multi-source data characteristics of chemical pharmaceutical due diligence reports create multiple constraints for the parsing and chunking process. If associated data in structured tables is split, the binding relationship between test data and corresponding batches will be lost. The parsing process must therefore retain row and column context for tables.
Long synthesis process description text is lengthy. Chunking operations must not interrupt the logical continuity of process steps. The binding relationship between professional fields and units must be retained to avoid mismatches between parameters and units during retrieval.
Documents with different update rhythms have wide span variations, from single-page raw material test reports to hundred-page clinical submission materials. This requires flexible adjustment of chunking granularity capabilities.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the length of long process descriptions and table blocks in chemical pharmaceutical documents, retaining complete context |
chunk_overlap | 150–200 characters | Retains associated information between table rows and the front/back logic of process steps |
PARSE_TABLE_STRUCTURE | Enabled | Accurately identifies structured tables in documents, retaining the binding relationship between test data and batch parameters |
UPLOAD_FILE_MAX_SIZE | 2000 MB | Supports uploading hundred-page clinical submission materials and patent literature |
PARSE_FILE_TIMEOUT_SECONDS | 1200 seconds | Adapts to parsing time for large documents, avoiding mid-process interruptions |
ENABLE_CHUNK_ANCHOR | Enabled | Uses unique identifiers such as CAS numbers and batch numbers as chunk anchors, improving retrieval accuracy |
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 specific analysis, and it is recommended to test on your own samples before finalizing settings.
Three common errors
- Phenomenon: After uploading chemical pharmaceutical due diligence documents larger than 10 MB, the generated chunks fail vectorization, and the log shows the
vectorization_errorstatus code. Cause: Thechunk_sizeparameter was not adjusted for long documents, causing single chunk content to exceed the context limit of the vectorization model. - Phenomenon: For multi-table documents imported via webhook, some table fields are empty after parsing. Cause: The
PARSE_TABLE_STRUCTUREconfiguration was not enabled, so column association relationships for multiple tables cannot be correctly identified. - Phenomenon: When attempting to import Feishu online spreadsheets into the knowledge base, only headers are retained without content after parsing. Cause: Feishu document authorization binding was not completed, and the authorization configuration for corresponding web spreadsheet parsing was not enabled, so complete spreadsheet content cannot be pulled.
How to confirm the configuration is correct
- Upload a single clinical submission document larger than 10 MB, check the number of parsed chunks and the length of each chunk, adjust
chunk_sizeuntil a single chunk covers a complete process paragraph or table row group. - Upload a raw material test report containing CAS numbers and purity parameters, check if the parsed chunks retain the association between fields and units, verifying the effect of the
ENABLE_CHUNK_ANCHORconfiguration. - Export the knowledge base's dataset.csv, confirm that the exported content includes chunked text from the original document and content beyond the template format, and check the status of the corresponding export configuration.
- Trigger a vectorization task, check that there are no
vectorization_errorstatus codes in the log, confirming that the chunking parameters adapt to the context limit of the vectorization model.
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.