Workflow Orchestration for Preclinical Safety Assessment and

Data Characteristics in Preclinical Safety Assessment

Preclinical safety assessment data originates from in vitro and in vivo (rodent and non-rodent) animal study reports, toxicology study reports, pharmacokinetic data, and pathology analysis results. This data typically exists in a mixed format, including structured data (e.g., CSV, Excel spreadsheets) and unstructured data (e.g., PDF research reports, pathology slide images, experimental record texts). Data update frequency is relatively low, usually occurring every few weeks or months as experimental batches and research phases complete. Document structures vary; for instance, toxicology reports include dose groups, animal counts, observation indicators (body weight, organ coefficients, blood biochemical indicators), statistical analysis results, and conclusions. Field names and units may differ slightly across laboratories or research phases. For example, dose might be expressed in mg/kg or μg/mL, and time point might be h, d, or w.

Constraints Imposed by These Characteristics on Workflow Orchestration

The mixed nature of preclinical safety assessment data requires workflows to support multimodal data processing, ingesting both structured and unstructured data. The low data update frequency means workflow triggers do not need to be overly frequent; on-demand or scheduled triggers are sufficient. The diverse document structures, especially inconsistent field names and units, necessitate strict standardization and normalization during data ingestion. For example, a unit conversion node might unify dose units to mg/kg. Large text volumes in unstructured reports demand higher chunk size and recall count for text embedding and retrieval to ensure no critical information is missed. Furthermore, the accumulation and referencing of historical data require workflows to support precise retrieval and comparison of specific batches or research reports, influencing knowledge base construction strategies and retrieval logic, potentially requiring metadata filtering.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`chunk size`	`500-800 characters`	Ensures the integrity of critical information sections in unstructured reports, preventing semantic fragmentation.
`recall count`	`10-20 items`	Balances retrieval efficiency with coverage, ensuring enough relevant toxicology data is recalled.
`similarity threshold`	`0.75-0.85`	Filters out low-relevance information, focusing on toxicology evidence highly matching prescreening criteria.
`rerank return count`	`5 items`	Reranks recall results to highlight the most relevant potential risks or safety data.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Handles the parsing requirements of large PDF research reports, preventing processing failures due to timeouts.
`maxContext`	`8000 Tokens`	Ensures the model can process the full context, including summaries and analysis results from multiple reports.

Three Common Pitfalls

During workflow execution, key fields in some reports are empty. This occurs when unit normalization and field mapping are not performed during data preprocessing, leading to the incorrect identification of synonymous fields from different sources.
The workflow encounters an HTTP 400 error when calling an external plugin for toxicity prediction. This happens when the global variable API_KEY is not configured correctly, leading to external service authentication failure.
Critical toxicity data is missing from preclinical safety assessment prescreening results. This is due to an excessively small chunk size during knowledge base construction, causing key toxicological conclusions in lengthy reports to be fragmented and not fully retrieved.

How to Verify Correct Configuration

Execute a workflow containing various data sources. Check the structured data output from each data ingestion node to ensure all key fields (e.g., dose, observation indicators, statistical results) are correctly parsed and units are consistent.
Set up test cases in the workflow. Simulate new compound information with known toxicity characteristics. Run the prescreening process and manually verify whether the output accurately identifies these known characteristics. Adjust the similarity threshold if necessary.
Check workflow logs. Confirm that no PARSE_FILE_TIMEOUT_SECONDS related timeout errors occur when processing large unstructured reports, and that text embedding and retrieval steps complete normally.
Verify the workflow's notification or report generation steps. Ensure that alerts are correctly triggered or detailed analysis reports are generated under predefined exceptional conditions (e.g., detection of high-risk signals).

The values provided are common starting points and should be measured against the reader's 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.