Data Characteristics in this Category
Stability study data primarily originates from drug production batch records, test reports, and related research protocols. This data updates infrequently, typically aligning with batch production and stability observation cycles. Document structures mix structured tables and unstructured text. They include batch information, observation conditions (temperature, humidity, light), test items (content, dissolution, degradation products), test results (values, units, acceptance criteria), and conclusions. Fields cover batch number, sample ID, sampling time point, test method ID, test value, and allowable range. Units vary significantly across different test items, including percentages, mg/tablet, pH values, and ppm.
Constraints on Multi-turn Conversations and Prompts
The low update frequency of stability study data means knowledge base synchronization can use periodic full or incremental updates, eliminating the need for real-time synchronization. The mixed structured and unstructured document format requires the Retrieval Augmented Generation (RAG) process to effectively handle both tabular data and natural language descriptions during retrieval, for example, through table extraction or multimodal embedding. Diverse fields and units demand more sophisticated prompt construction. The model must recognize and correctly interpret numerical values with different units to avoid confusion. For instance, when a user asks about "degradation product content," the model needs context to distinguish between a percentage and an absolute quantity. Additionally, large volumes of historical batch data necessitate effective context management in multi-turn conversations. This ensures the model can track user queries that jump between different batches and test items, maintaining conversational coherence.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 500–800 characters (characters) | Stability reports often have long paragraphs containing multiple test items or batch information. Extending the segment length helps maintain contextual integrity. |
Recall count (Recall Count) | 8–12 entries (items) | Stability research questions typically involve multiple related batches or test standards. Increasing the recall count improves the probability of retrieving relevant information. |
Similarity threshold (Similarity Threshold) | 0.75 | Policy documents are rigorous, requiring a high similarity threshold to avoid recalling irrelevant clauses. |
maxContext | 3000–4000 Token | In multi-turn conversations, users may frequently switch between different batches or test points. A longer context window is needed to maintain conversational coherence. |
Rerank result count (Reranked Return Count) | 5 entries (items) | This improves the precision of recall results, prioritizing the most relevant pieces of information for the model. |
Three Common Mistakes
- Symptom: Model responses contain incorrect or missing numerical units. Reason: Prompts do not explicitly instruct the model to pay attention to and output numerical units, or relevant fields in the knowledge base are not standardized for units.
- Symptom: In multi-turn conversations, the model fails to correctly understand user queries that switch between different batch data. Reason: The
maxContextparameter is set too low, causing batch information from earlier conversation turns to be lost, and the model cannot track the context. - Symptom: When a user queries a specific test item, the model recalls many irrelevant policy clauses. Reason: The knowledge base segmentation strategy is too coarse, failing to effectively differentiate detailed regulations for different test items, leading to an overly broad recall range during similarity matching.
How to Verify Configuration
- Conduct multi-turn conversation tests for typical stability research questions. Check the model's ability to track key information like batches, test items, and time points across different turns.
- Randomly select critical data points from stability research reports. Query the model and verify the accuracy of numerical values, units, and acceptance criteria in the model's responses.
- Simulate a user asking about a specific clause in the policy. Observe the knowledge items recalled by the model. Confirm the high relevance of the recalled content to the question and check for any unnecessary redundant information.
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.