Data Characteristics
Patient Assistance Program (PAP) quality documentation data primarily originates from pharmaceutical companies' regulatory, medical affairs, and market access departments, as well as partner charitable organizations. These documents typically include program policies, application guidelines, approval processes, patient informed consent forms, drug dispensing records, adverse event report templates, and audit records. Document update frequencies vary. Policy documents may be revised semi-annually or annually. Patient application forms or dispensing records update in real time.
Document structure is mostly unstructured text, such as Word and PDF policy files. It also includes structured data, such as patient information and medication records in Excel or databases. Fields and units are highly specific. Examples include drug batch numbers, expiry dates, patient medical record numbers, diagnostic codes (e.g., ICD-10), dosage (mg/kg), assistance period (months/years), and approval status (pending, approved, rejected).
Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts
The unstructured text nature of patient assistance documents, especially policy and guideline files, requires accurate extraction of key information in multi-turn conversations. This includes assistance criteria, application material checklists, and specific disease medication regulations.
Inconsistent document update frequencies, particularly real-time patient records, mean the system needs fast indexing and knowledge base update capabilities. This ensures the timeliness of conversation content.
Highly specific fields and units, such as drug dosage and diagnostic codes, require prompt design to guide the model in understanding and differentiating these specialized terms. This avoids confusion.
These documents often involve legal regulations and privacy protection. The conversation system must strictly adhere to data isolation and access control. This ensures no sensitive information is leaked during multi-turn interactions. It also ensures correct citation of original document sources to enhance answer credibility and compliance.
Configuration Strategy
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 500-800 characters | Ensures semantic completeness of policy and guideline paragraphs, preventing truncation of key information. |
Recall count (Recall Count) | 8-12 items | Covers multiple policy clauses and related attachments, improving recall comprehensiveness. |
Similarity threshold (Similarity Threshold) | 0.75-0.85 | Balances recall precision and recall rate. Avoids irrelevant clauses while not missing key information. |
Rerank result count (Reranked Return Count) | 3-5 items | Focuses on the most relevant policies or process details, optimizing the final answer quality for the user. |
maxContext | 4096 tokens | Accommodates complex policy interpretation and multi-turn follow-up scenarios, allowing for longer conversation history and recalled content. |
System Prompt | Includes "As A Patient Assistance Program Expert" (As a patient assistance program expert) and "Answer Only Based On The Provided Document" (Answer only based on the provided documents) | Defines the role and forces the model to provide information within compliance boundaries, avoiding hallucinations. |
Common Pitfalls
- The model suddenly fails to answer questions about specific clauses in a multi-turn conversation. This may be due to an unreasonable document segmentation strategy, leading to key information being split or lost, affecting recall completeness.
- When users ask about specific drug dosages or diagnostic codes, the returned answers lack professionalism or contain unit errors. This usually happens when the prompt fails to adequately emphasize the requirements for identifying and processing specialized fields.
- LaTeX formulas display correctly in the debugging preview but appear as raw code in the chat box after publication. This indicates the front-end rendering component is not configured correctly or lacks a LaTeX parsing library.
Verification Steps
- Select at least 10 typical patient assistance policy documents. Ask questions about key clauses, application procedures, and contraindications. Check if answers accurately cite the original text and handle multi-turn follow-ups correctly.
- Input queries containing professional terms and units (e.g., "10mg/kg" or "ICD-10 code"). Verify if the model correctly identifies and explains this information. Also, confirm that related answers do not contain unit or numerical errors.
- Simulate user inquiries about project update frequency. Check if the system can indicate the latest document update date or point out which documents are updated in real time. This verifies the data timeliness mechanism.
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.