Off-label Use Medical Information (MI) Response with

Data Characteristics for This Category

Off-label use medical information data primarily originates from professional medical journals, clinical guidelines, announcements from drug regulatory agencies, expert consensus, and real-world evidence (RWE). Data update frequencies vary but are generally high, especially when new indications or safety information are involved. Document structures are complex and may include research papers, clinical trial reports, case reports, and reviews. These documents cover drug indications, dosages, pharmacological mechanisms, adverse reactions, and contraindications. A specific characteristic is the frequent inclusion of various disease classification codes (e.g., ICD-10), drug classification codes (e.g., ATC), and dosage units (mg/kg, U/mL), as well as time units (hours, days, weeks). The standardization of these fields and units varies significantly, and they are often accompanied by extensive unstructured descriptions.

Constraints Imposed by These Characteristics on "Multi-turn Conversations and Prompts"

The complexity of off-label use data directly impacts the design of multi-turn conversations and prompts. First, diverse data sources and frequent updates require the knowledge base to have efficient indexing and real-time synchronization mechanisms to ensure the accuracy and timeliness of conversational content. Second, complex document structures and unstructured descriptions necessitate more refined instructions in prompt engineering to guide the model in extracting key information from lengthy texts. For example, when a user asks about adverse reactions of a specific drug in an off-label use scenario, the model must accurately identify and summarize relevant information scattered across multiple documents. The presence of disease and drug classification codes requires prompts to recognize and utilize these professional terms for precise retrieval and contextual association. Furthermore, the diversity of dosages and units means that understanding and generating medication plans in multi-turn conversations requires additional attention to unit consistency and numerical reasonableness to prevent misunderstandings or incorrect advice.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8192 token`	Ensures the model can handle complex contexts, including summaries of multiple documents and conversation history, especially the lengthy background information in off-label use scenarios.
`Chunk size` (Segment Length)	`500–800 characters`	Accommodates the characteristics of medical literature, which often has long paragraphs and high information density, balancing recall accuracy with processing efficiency.
`Recall count` (Recall Items)	`8–12 entries`	Increases the probability of retrieving relevant evidence from a vast amount of off-label use data, covering information from different sources and perspectives.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Filters out low-relevance documents, ensuring retrieved medical information highly matches user queries and reduces noise.
`Rerank result count` (Reranked Return Items)	`4–6 entries`	Further refines the most relevant evidence based on initial recall, improving the accuracy and focus of multi-turn conversations.
`temperature`	`0.3–0.5`	Reduces the randomness of the model's generated content, ensuring the rigor and objectivity of medical information responses, and avoiding speculation.

Three Common Mistakes

Frontend requests to the chat interface http://localhost:3000/api/v1/chat/completions result in a CORS error: This usually occurs when the frontend application and FastGPT backend are deployed on different domains, and the backend has not correctly configured the Access-Control-Allow-Origin header.
An HTTP request configured in the workflow is not triggered, and an AI conversation occurs directly: This is often due to imprecise conditional logic in the workflow, causing the trigger conditions for the HTTP request branch to not be met, or the model generating an answer without explicit instructions.
In a multi-turn conversation, the response from the first AI conversation is used as input for the second AI conversation, but the final result still includes content from the first AI conversation: This typically happens because the prompt for the second AI conversation does not explicitly instruct it to output only new content, or the workflow's output configuration does not exclude intermediate step results.

How to Confirm Proper Configuration

Simulate multi-turn conversations for typical off-label use queries, and cross-reference the model's output for drug dosages, indications, and adverse reactions with original medical literature for consistency.
Randomly select a number of queries and check if HTTP requests in the workflow are triggered as expected, verifying that their returned data is correctly parsed and utilized.
In complex multi-turn conversation scenarios, review the output of each AI conversation step to confirm the accuracy of intermediate results and verify that the final output contains only the answers relevant to the user.
Use logs or debugging tools to confirm the accuracy of knowledge base recall and model understanding when processing queries containing professional terms (e.g., ICD-10 codes).

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.