Multi-Turn Conversations and Prompts for Pharmacovigilance

Stability Study Data Characteristics

Stability study data originates from long-term, accelerated, and intermediate study reports during drug development. These reports are typically PDF or Word documents. They contain detailed batch information, observation time points, temperature/humidity conditions, test items, test results, and judgment criteria. Data updates are infrequent, usually completed before drug commercialization. Subsequent updates occur only with significant changes or special circumstances. Document structures typically include a preface, study protocol, study results (often in tabular form with specific fields like Batch Number, Observation Time, Temperature, Humidity, Assay, Related Substances, pH Value), data analysis, and conclusions. Assay units are commonly % or mg/g, related substances units are %, and pH values are unitless.

Constraints on Multi-Turn Conversations and Prompts from These Characteristics

The documented nature of stability study data requires multi-turn conversations to focus on precise extraction and understanding of structured and semi-structured information. Due to the low frequency of data updates, knowledge base construction and maintenance can use periodic full updates. The Recall Count setting must cover sufficient time points and batch information. In multi-turn conversations, users may ask follow-up questions about specific batch trends over different time points or assay degradation details under specific conditions. This requires the system to accurately link relevant data across different documents. Additionally, specialized terminology and units in the data (e.g., Degradation Product, Peak Area, Relative Standard Deviation) demand higher precision in Prompt design. This ensures the model correctly identifies and interprets this specialized content.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk Length`	`500–800 characters`	In stability reports, individual test results or batch information paragraphs are of moderate length. Excessive length can introduce irrelevant information, while insufficient length may break critical context.
`Recall Count`	`Top 10–15 entries`	Ensures that multi-turn conversations cover multiple batches, time points, or test items that users might trace, preventing information omission.
`Similarity Threshold`	`0.75–0.85`	Stability study terminology is highly specialized. Increasing the threshold reduces the recall of irrelevant segments, improving answer accuracy.
`Rerank Return Count`	`Top 5 entries`	Further refines the most relevant snippets from the recalled results, optimizing the quality and conciseness of the final answer.
`maxContext`	`4096 tokens`	Provides the model with sufficient context to handle complex user queries in multi-turn conversations and link multiple stability data points.
`System Prompt Version`	`V4.9.13`	Ensures the use of the latest model capabilities and prompt parsing logic for better understanding and generation.

Common Pitfalls

Symptom: The user asks about the assay of a specific batch at a particular time point. The AI responds with "no relevant information found" or provides incorrect data. Cause: Inadequate knowledge base chunking strategy leads to critical data (e.g., Assay field) being split across different segments, or the Similarity Threshold is too high, filtering out relevant segments with slightly lower similarity.
Symptom: In a multi-turn conversation, the user asks a follow-up question about a Related Substance mentioned in the previous AI response, but the AI cannot link the context. Cause: The maxContext parameter is set too low, causing historical conversation information to be truncated in subsequent turns, preventing the model from accessing the complete dialogue history.
Symptom: The AI confuses Observation Time or Batch Number in its response, leading to incorrect data correlation. Cause: The Prompt does not explicitly instruct the model to strictly differentiate and cite specific fields from the document in its answer, or the Recall Count is insufficient, preventing the model from obtaining enough information to distinguish between different batches or time points.

Verification Steps

Conduct multi-turn tests for typical stability study questions (e.g., "What is the trend of Assay for Batch XYZ at 36 months under 40℃/75%RH conditions?"). Check if the AI can accurately trace data, link context, and provide reasonable explanations.
Verify if the AI's response accurately cites key fields from the document, such as Batch Number, Observation Time, and Test Item, ensuring data traceability.
In the FastGPT backend, review the actual recalled content corresponding to Recall Count and Similarity Threshold. Determine if the recalled segments contain all critical information required for the user's query and exclude irrelevant information.
Simulate long conversation tests with different maxContext settings. Observe the AI's ability to remember and link early information in the later stages of the conversation.

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.