Multi-Turn Conversations and Prompts for DTP Pharmacy

Data Characteristics for This Category

Data for Direct To Patient (DTP) pharmacy registration document preparation primarily comes from official drug administration websites, pharmaceutical company product inserts, clinical trial reports, drug registration approvals, and the pharmacy's operational data. This data updates infrequently, typically with new drug approvals or policy changes. Document structures are mainly PDFs, Word files, and scanned images, containing significant unstructured text and tables. Specific fields include drug indications, dosage and administration, adverse reactions, contraindications, as well as pharmacy GSP certification, cold chain logistics qualifications, and pharmacist credentials. Units commonly include milligrams (mg) and milliliters (ml) for dosage, and days, months, and years for time.

Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts

The varied document formats and large amount of unstructured content in DTP pharmacy data require the multi-turn conversation system to have robust document parsing capabilities for knowledge base construction, especially for accurate extraction of tables and text from PDFs and scanned images. The low update frequency means the knowledge base can undergo periodic batch updates, but each update must ensure completeness and consistency. Unique medical terminology and specialized fields in documents require prompt design to consider terminology accuracy and contextual relevance, preventing misinterpretation due to misunderstandings. Furthermore, DTP pharmacy operational qualifications are often critical for compliance review. Multi-turn conversations must accurately locate and cite relevant qualification documents, demanding higher precision in prompt recall.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	Ensures each segment contains sufficient contextual information while avoiding excessive length that could lead to semantic drift, especially for specialized medical documents.
`Recall count` (Recall Count)	`Top 8`	Given the complexity and interconnectedness of DTP pharmacy data, increasing the recall count helps cover more potentially relevant information.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	The medical field demands high information accuracy; a strict threshold avoids recalling irrelevant or ambiguous content.
`maxContext`	`8192 tokens`	Ensures multi-turn conversations can handle sufficiently long contexts to address complex registration questions and trace historical dialogue.
`Rerank result count` (Reranked Return Count)	`Top 3`	Performs a secondary sort on recalled results, prioritizing the most relevant core information for improved efficiency.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	DTP pharmacy documents may contain many images and complex tables, requiring a longer timeout for parsing.

Three Common Mistakes

Long delays between data responses in conversations lead to a poor frontend user experience. This occurs when chunk size or transmission rates for streaming output are not optimized, resulting in infrequent data packet delivery.
Multi-turn conversations cite irrelevant drug approval information, causing discrepancies in registration documents. This happens when the similarity threshold is set too low or knowledge base segmentation granularity is too large, leading to the recall of similar content from non-target drugs.
Missing critical interaction records in conversation logs prevent problem tracing. This is due to improper logging level or storage policy configuration, failing to capture all user inputs and system outputs completely.

How to Confirm Proper Configuration

Submit complex questions involving medical terminology and qualification requirements. Observe if the system provides highly relevant responses within 5 seconds and cites correct document snippets.
Upload a DTP pharmacy GSP certification document containing tables and scanned text. Check if knowledge base segmentation accurately identifies all key fields, such as certificate_no and valid_until.
Simulate multi-turn follow-up questions from a user, for example, first asking about "a drug's indications," then "that drug's contraindications." Verify if the system maintains contextual relevance for the drug information in subsequent conversations.
Check if conversation logs completely record each user query, system retrieval process, and final response content, ensuring continuity of dialog_id and turn_id fields.

These values are common starting points and should be measured against your 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.