Multi-turn Conversation and Prompt Design for Rational Drug

Data Characteristics for This Category

Data for rational drug use in special populations comes from authoritative medical guidelines, drug inserts, pharmacopoeias, clinical research reports, and consensus documents for specific groups (e.g., pregnant women, lactating women, children, the elderly, patients with hepatic or renal impairment). This data exists as unstructured text, semi-structured tables, and structured databases. Update frequency varies: drug inserts and pharmacopoeias are typically revised annually or irregularly, while clinical guidelines and research reports may have longer update cycles, usually 2–5 years. Document structures are complex, containing fields such as dosage, administration, contraindications, precautions, drug interactions, and adverse reactions. Units include mg, ml, kg, times/day, μg/kg/min, often accompanied by ranges or conditional limitations.

Constraints Imposed by These Characteristics on "Multi-turn Conversation and Prompt Design"

The authoritative nature of the data demands accuracy and rigor in question-answering results. Prompt design must emphasize extracting key information from original texts and avoiding generative hallucinations. Differing update frequencies necessitate sophisticated version management and incremental update mechanisms for the knowledge base, ensuring multi-turn conversations are based on the latest knowledge. Complex document structures require segmentation strategies that effectively identify and preserve the integrity of critical information blocks, such as drug contraindications or specific dosage adjustment rules. The variety of fields, units, ranges, and conditional limitations challenges prompt instruction precision. Prompts need to clearly guide the model on how to parse and present these values and restrictions, ensuring accurate quantitative comparisons and conditional judgments in multi-turn conversations. For example, when asked "Can a pregnant woman use this drug?", the model must accurately identify specific descriptions regarding drug use in pregnant women from the drug insert.

Configuration Settings

Configuration Item	Suggested Value	Rationale for This Value
`maxContext`	`6` turns	Balances context continuity with computational resources, covering typical rational drug use inquiry scenarios.
`Chunk size` (Segment Length)	`800–1200 characters`	Preserves the integrity of pharmaceutical knowledge blocks, preventing truncation of critical information.
`Recall count` (Recall Count)	`Top 5–8 entries`	Increases coverage of relevant knowledge points, addressing complex queries and multi-condition judgments.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Ensures precision of recalled content, reducing interference from irrelevant information.
`Rerank result count` (Rerank Return Count)	`Top 3 entries`	Highlights the most relevant knowledge snippets, optimizing response efficiency in multi-turn conversations.
`promptTemplate`	Include "Strictly answer based on provided medical literature; do not speculate."	Emphasizes rigor and fact-based answers, preventing information bias.

Three Common Mistakes

The knowledge base fails to correctly parse document content after file upload via the conversation interface, preventing the system from referencing file information in multi-turn conversations. This results in empty or generalized responses. This occurs because PARSE_FILE_TIMEOUT_SECONDS is set too short or the file type does not match the parser.
Knowledge base reference prompts are not translated into English, causing the model to be unable to effectively recall or understand English queries. This typically happens when the promptTemplate parameter is not configured for a multi-language environment or an English version of the prompt is not provided.
Context memory is interrupted in continuous conversations, leading users to repeatedly mention previously discussed information in subsequent questions. This manifests when the maxContext parameter is set too low, truncating historical conversation turns and failing to effectively pass context.

How to Verify Correct Configuration

Ask multi-turn questions about drug use for various special populations (e.g., pregnant women, children, the elderly). Check if each answer accurately references relevant entries in the knowledge base.
Upload a drug insert containing complex tables and descriptions with multiple units. Verify if the knowledge base can correctly parse it and extract key values like dosage and administration in question-answering.
Simulate user queries in different language environments (e.g., English). Confirm if the system can correctly understand and provide relevant answers, and check if promptTemplate includes corresponding multi-language prompts.
Check the actual effective value of the maxContext parameter in the logs and compare it with the actual conversation turns to confirm context is passed as expected.

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.