Data Characteristics
Patient Assistance Program (PAP) pharmacovigilance data originates from voluntary patient reports, healthcare institution submissions, and program management collection. Data update frequency is irregular, typically aggregated monthly or quarterly, with urgent adverse event reports updated in real-time. Document structure primarily consists of structured tabular data, including fields such as patient basic information, medication history, adverse event description, treatment measures, severity, and event outcome. Some data may exist as unstructured text, such as patient interview records or physician diagnostic notes. Field units involve dosage (e.g., milligrams, grams), frequency (e.g., once daily, twice weekly), and time (e.g., hours, days, weeks). Data characteristics include high sensitivity, involvement of personal privacy, and potential subjectivity and ambiguity in adverse event descriptions.
Constraints Imposed by These Characteristics on "Citation and Traceability"
High sensitivity requires strict access control for citations, ensuring data security and compliance. The presence of unstructured text increases the difficulty of direct citation, demanding more refined text segmentation and semantic understanding capabilities. Irregular data update frequency means knowledge base content needs regular synchronization to avoid citing outdated information. The subjectivity and ambiguity of adverse event descriptions can lead to inaccurate recall during similarity matching, affecting traceability precision. Furthermore, aggregating multi-source data makes it difficult for a single citation link to fully present the original context. This necessitates supporting the aggregated display of multiple citation paths to fully reconstruct the reporting background and treatment process of adverse events.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size | 500–800 characters | Balances the completeness of adverse event descriptions with recall efficiency, preventing semantic breaks or information redundancy from overly long or short segments. |
Recall count | Top 8 entries | Covers potentially relevant adverse event reports or patient records, increasing the comprehensiveness of traceability. |
Similarity threshold | 0.75 | Balances recall precision and coverage, reducing misjudgments caused by subjective descriptions. |
maxContext | 3000 Tokens | Ensures sufficient context retention in multi-turn conversations, especially for follow-up questions regarding complex adverse events. |
citation Display Mode | Paragraph Add Link To End | Complies with regulatory requirements, facilitating quick navigation to original reports or records. |
Sensitive Data Anonymization | Enabled | Ensures patient privacy protection by automatically processing personal identification information during citation display. |
Three Common Mistakes
- Phenomenon: Knowledge base answers do not display citation or "view original" functionality. Reason: The "Citation Source" function is not enabled in the application configuration, or permission settings for password-free sharing links restrict the display of this function.
- Phenomenon: The content cited in the answer does not fully correspond to the actual original paragraph. Reason: The text segmentation strategy is too coarse, causing a single segment to contain too much irrelevant information or critical information to be split across different paragraphs.
- Phenomenon: During continuous follow-up questions, the answer deviates from the topic and lacks contextual connection. Reason: The
maxContextparameter is set too small, preventing the model from effectively utilizing previous conversation history in multi-turn dialogues.
How to Confirm Correct Configuration
- Simulate multi-turn conversations, focusing on details such as adverse event severity and treatment processes. Check whether each answer displays corresponding citation sources and allows clicking to navigate to key locations in the original document.
- Randomly select multiple adverse event reports. Use their key information as queries to check whether the recalled citation sources are comprehensive and relevant, especially whether critical patient reports or medical records are missing.
- Use test data containing sensitive information for queries. Confirm that personal identification information (e.g., name, ID number) is effectively desensitized during citation display.
The values given 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.