Multi-turn Conversations and Prompts for Stability Study

Data Characteristics in Stability Studies

Stability study data originates primarily from experimental reports, batch production records, and analytical method validation documents. Data updates typically align with product batch production or annual reviews, showing periodic updates, such as large-volume data entry every six months or annually. Document structures are often structured or semi-structured. Examples include tabular data for physicochemical index trends, microbial test results, and assay data. Key metadata like batch number, production date, expiration date, and storage conditions accompany this data. Fields include, but are not limited to: test item, test result, unit (unit, e.g., ug/mL, %, CFU/g), test time point, standard range, and deviation. Some documents may include scanned images or graphs, requiring specialized handling for these non-textual elements.

Constraints from These Characteristics on Multi-turn Conversations and Prompts

The periodic update cycle of stability study data requires knowledge base updates to occur in batches or annually, preventing the introduction of outdated information. The high degree of document structure allows for more precise specification of fields to extract in prompt design. For example, a prompt can directly request assay results for a specific batch at a particular time point. The specificity of units requires the model to correctly identify and output them, avoiding confusion between ug/mL and %. The presence of non-textual information, such as graphs, means that if a user mentions "abnormal graph" in a multi-turn conversation, the AI system needs to prompt the user to upload an image or provide an image description. Version v4.8.10 processes text by default; image processing requires additional extensions. Furthermore, the large volume and high repetitiveness of the data place high demands on the settings for Recall count (number of recalled items) and Similarity threshold (similarity threshold), requiring a balance between recall accuracy and response speed.

Configuration Settings

Configuration Item	Suggested Value	Rationale for this Value
`maxContext`	`8000 tokens`	Stability reports are often long, requiring a larger context window to accommodate historical information and referenced document content in multi-turn conversations.
`Chunk size` (Chunk Length)	`500 characters` (characters)	This ensures each knowledge chunk contains sufficient context while avoiding excessive length that could lead to information redundancy or inaccurate segmentation.
`Recall count` (Number of Recalled Items)	`Top 8 entries` (top 8)	Stability data has strong interconnections. Increasing the number of recalled items enhances the comprehensiveness of retrieved relevant batches and test items.
`Similarity threshold` (Similarity Threshold)	`0.78`	Stability data fields are named consistently. A higher threshold reduces the erroneous recall of irrelevant batches or test items.
`Rerank result count` (Number of Reranked Items)	`Top 3 entries` (top 3)	After initial recall, the most relevant few items are selected for display, improving the accuracy and conciseness of the response.
`Temperature`	`0.3`	Stability studies require objective facts. A lower temperature reduces the model's free-form generation, ensuring the accuracy of the response.

Three Common Pitfalls

Symptom: When calling the conversation API, log details and actual response content do not match. Repeated queries still show the same old data. Reason: The knowledge base does not dynamically update its retrieval strategy based on the latest session state during multi-turn conversations, or caching mechanisms prevent timely refreshing of relevant document segments.
Symptom: In multi-turn conversations, the model fails to correctly understand user queries involving specific units (e.g., nM or IU/mg), leading to unit confusion or missing information in the response. Reason: The prompt does not explicitly instruct the model to pay attention to unit information, or unit fields were not standardized during knowledge base data entry.
Symptom: A user asks about the microbial limit result for a specific batch, but the system returns assay data, or states "no relevant information found." Reason: The Similarity threshold (similarity threshold) is set too high, preventing relevant documents from being recalled; or the Chunk size (chunk length) is too small, causing critical information to be split into different segments and making it difficult to identify.

How to Verify Correct Configuration

Conduct multi-turn conversation tests for different batches and different test items (e.g., dissolution, related substances). Check if the system can accurately extract and present the corresponding data, verifying the effectiveness of Similarity threshold (similarity threshold) and Recall count (number of recalled items).
Input queries containing specialized units (e.g., ug/mL, CFU/g). Observe if the model's output correctly references or converts units, confirming the prompt's guidance for unit recognition.
Simulate users asking follow-up questions about historical data or comparing data from different batches. Check if the model can maintain contextual coherence and adjust the retrieval scope based on historical conversation content, verifying the effect of the maxContext configuration.

Note: The values provided are common starting points and should be measured against specific 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.