Preclinical Safety Assessment Pharmacovigilance: Multi-turn

Data Characteristics

Preclinical safety assessment data originates primarily from toxicology reports, pharmacokinetic reports, pathology analyses, and relevant regulatory documents. This data typically exists as unstructured or semi-structured documents, such as detailed research reports in Word or PDF format, or biomarker data in Excel spreadsheets. Update frequency is relatively low, with updates usually occurring in batches after completing specific research phases. Document content is extensive, including animal study protocols, dosage settings, observation indicators, and statistical analysis results. Fields and units are diverse, covering physiological indicators (e.g., weight, temperature, in grams, Celsius), blood biochemical indicators (e.g., liver enzyme activity, in U/L), pathological descriptions (e.g., tissue lesion severity, unitless or descriptive grading), and drug exposure (e.g., AUC, Cmax, in ng·h/mL).

Constraints Imposed by Data Characteristics on Multi-turn Conversations and Prompts

The detailed and specialized nature of preclinical safety assessment reports requires the conversational system to deeply understand context and accurately extract multi-level information. Due to the low data update frequency, the knowledge base content remains relatively stable, allowing for deep indexing of historical data. However, in multi-turn conversations, users might inquire about differences between various experimental batches or species, necessitating cross-document retrieval capabilities. The extensive use of specialized terminology, abbreviations, and mixed units in documents challenges prompt robustness. Prompt design must effectively guide the model to identify and differentiate this information, preventing irrelevant responses due to unit confusion or misunderstood terminology. Furthermore, the complex report structure allows users to ask jump-around questions during follow-ups, for example, shifting from a specific toxicity indicator to its corresponding dosage group. This requires the multi-turn conversational system to flexibly switch between different information dimensions while maintaining conversational coherence.

Configuration Strategy

Configuration Item	Recommended Value	Rationale
`maxContext`	`8192 token`	Covers most follow-up scenarios, balancing cost and performance.
`Chunk size` (Segment Length)	`800 characters`	Ensures complete semantic units in each segment, avoiding cutting off key descriptions.
`Recall count` (Recall Count)	`10 entries`	Increases relevant information coverage for complex queries.
`Similarity threshold` (Similarity Threshold)	`0.78`	Balances recall rate and accuracy, reducing irrelevant information interference.
`Rerank result count` (Rerank Return Count)	`5 entries`	Prioritizes the most relevant content, reducing model processing load.
`Citation template` (Citation Template)	`Source: {{url}}`	Facilitates user traceability to original report sources.

Common Pitfalls

Symptom: The model starts "hallucinating" during the second follow-up question, unable to connect to the first turn of the conversation. Reason: The maxContext parameter is set too low, preventing the model from retaining enough historical conversation information, leading to context loss.
Symptom: Even when the knowledge base contains relevant content, the model fails to recall the correct document or recalls many irrelevant documents. Reason: The Similarity threshold (Similarity Threshold) is set incorrectly; too high leads to insufficient recall, too low leads to excessive noise.
Symptom: The model's response contains unit confusion or numerical errors, for example, misinterpreting mg/kg as g/kg. Reason: The prompt does not explicitly instruct the model to pay attention to units and dimensions, or key numerical values and units are separated during knowledge base document segmentation.

Validation of Configuration

Select multiple sets of preclinical safety assessment reports containing different toxicity indicators, dosage groups, and animal species. Conduct multi-turn questioning tests to observe if the model can continuously track the conversation topic.
For specific toxicity data in reports (e.g., LD50 or NOAEL values), verify if the model can accurately extract numerical values and units, and trace them back to the original report paragraphs using the Citation template (Citation Template).
Simulate user questions comparing and analyzing data from different experimental groups or time points. Check if the model can correctly understand and integrate cross-document information.
Examine if the model can maintain correct understanding and interpretation of specialized abbreviations or specific terminology (e.g., ALT, AST, AUC) in multi-turn conversations.

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.