Data Characteristics
Data for lead compound screening in pharmacovigilance originates from high-throughput screening reports, activity validation data, preliminary toxicity assessment reports, and compound structure databases. This data is typically structured (e.g., compound ID, activity values, toxicity indicators) and semi-structured (e.g., experimental method descriptions, batch information, result interpretations). Update frequency is irregular, usually generated after each screening batch or supplemented after further validation of active compounds. Document formats vary, including CSV, SDF, PDF reports, and XML or JSON files exported from internal LIMS systems. Key fields include SMILES strings, CAS numbers, IC50/EC50 values, cytotoxicity data (e.g., CC50), adverse reaction signals (e.g., hERG inhibition rate, CYP450 inhibition rate), and experimental conditions. Activity values are typically in nanomolar (nM) or micromolar (µM) units.
Constraints on "Reference and Traceability"
The diversity of lead compound screening data complicates reference source parsing, requiring support for multiple file formats. Non-structured experimental methods and result interpretations demand advanced text extraction capabilities for effective utilization. The uncertain update frequency necessitates flexible data ingestion mechanisms to avoid duplicate imports or missing new data. Compound structures, as core identifiers, require precise matching and retrieval; subtle differences in SMILES strings can lead to citation failures. Quantitative data for toxicity indicators and adverse reaction signals require accurate recognition of numerical ranges and units to ensure traceability. Additionally, the correlation between different data batches poses a challenge for building citation chains, requiring metadata management to establish clear associations.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size | 500–800 characters | Balances the completeness of structured data fields with the contextual coherence of non-structured descriptions, preventing key information from being truncated. |
Recall count | 8–12 entries | Balances retrieval efficiency with information coverage, ensuring recall of relevant lead compound data from various sources. |
Similarity threshold | 0.75–0.85 | Considers the text similarity of compound structures (SMILES) and experimental descriptions, avoiding over-recall or under-recall. |
Rerank result count | 3–5 entries | Focuses on the most relevant preliminary screening results and adverse reaction signals, improving the precision of final citations. |
Index Update Cycle | On-demand trigger，OrWeekly | Given the non-periodic updates of screening data, on-demand triggering promptly incorporates new data, with regular updates as a fallback. |
Parser Strategy | Multimodal Parsing（Text+Structured） | Supports simultaneous processing of text descriptions in PDF reports and structured activity data in CSV/SDF, ensuring comprehensiveness. |
Common Mistakes
- The page view model context displays 30 items, but 310 items are sent to OneAPI. This might be due to front-end display limitations or a different context truncation strategy.
- The citation limit is set to 1500 characters, but blocks exceeding 1500 characters in the knowledge base are still cited. This usually occurs because the segmentation strategy does not strictly adhere to character count, or the system's definition of "block" differs from user understanding.
- Compound IDs or SMILES strings fail to match during citation. This is because structured fields were not standardized during data import, or retrieval did not account for case sensitivity, isomers, or other subtle differences.
How to Verify Configuration
- Check system logs for the number of recalled items and the final context length passed to the large model for each query, confirming consistency with configured parameters.
- Select several representative lead compounds and query them. Verify that the returned reference sources include key documents like high-throughput screening reports and toxicity assessment data, and cross-reference the activity values and adverse reaction signals within them.
- Randomly sample some document blocks from the knowledge base that exceed the
Chunk size. Check if they are correctly segmented and verify that each segment can be independently recalled.
Note: The values provided are common starting points. Measure them against your own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.