Multiturn Conversation and Prompts for Dermatology R&D

Data Characteristics

Dermatology R&D documents include clinical trial protocols, case report forms (CRFs), investigator brochures (IBs), medical literature reviews, drug inserts, and patent applications. These documents originate from various sources, such as pharmaceutical company internal reports, academic journals, and regulatory approvals. Update frequency varies by document type; clinical trial data and literature reviews may update frequently, while drug inserts and patent files are relatively stable. Document structures are typically highly standardized. For example, CRFs follow the CDISC SDTM/ADaM model and contain many structured fields. Field content covers patient demographic information, disease diagnosis, treatment plans, drug dosages, adverse event (AE) and serious adverse event (SAE) records, and laboratory test results. Units commonly include dosage units (mg, g), time units (days, weeks), area units (cm²), and various biological indicator units (e.g., IU/L, μg/mL).

Constraints Imposed by These Characteristics on Multiturn Conversation and Prompts

The standardized structure and rich fields of dermatology R&D documents require the multiturn conversation system to have precise entity recognition and relationship extraction capabilities. For example, when processing CRFs, the system must accurately identify key fields like "patient ID," "drug dosage," and "adverse event type," and understand their relationships. Frequently updated clinical data and literature mean the knowledge base needs to support efficient incremental updates and version management to ensure multiturn conversations are based on the latest information. The specialized terminology and abbreviations within fields (e.g., "AD" for atopic dermatitis, "BSA" for body surface area) require prompt design to consider terminology standardization and synonym mapping to avoid retrieval failures due to inconsistent terms. Additionally, numerical data involving drug dosages and laboratory results requires the conversation system to perform numerical comparisons and range queries, such as "find all patients with BSA greater than 1.5 m²."

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`maxContext`	`8`	Dermatology R&D queries often require a longer context to understand complex medical histories and treatment plans.
`Chunk size`	`500–700 characters`	Ensures each text block contains sufficient information while avoiding excessive length that could dilute semantics, suitable for detailed clinical descriptions.
`Recall count`	`Top 8 entries`	Considering document complexity and cross-references, increasing the number of recalled items improves relevant information coverage.
`Similarity threshold`	`0.78–0.85`	Dermatology terminology demands high precision; a threshold that is too low may introduce noise, while one that is too high may miss relevant results.
`Rerank result count`	`Top 3 entries`	After reranking, selecting a small number of the most relevant results reduces the model's processing burden and improves response speed.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Addresses cases where parsing large clinical trial reports or review documents takes a long time.

Common Pitfalls

Uploading large documents results in a 503 Service Unavailable error, but backend logs show successful file upload. This typically indicates a file parsing timeout, where the processing result was not returned in time.
When referencing knowledge base image URLs in a multiturn conversation, the system replies "no answer found." This occurs when the strict Q&A template is not configured for image content recognition or image description extraction capabilities.
When questions are sent in rapid succession, subsequent questions are not processed immediately but wait for the previous request to complete. This may be due to limitations in the concurrent request handling mechanism or bottlenecks in model generation speed.

Verification Steps

Upload a clinical research report containing common dermatological conditions (e.g., psoriasis, eczema) and test whether the system can accurately parse patient baseline information, drug dosages, and adverse reactions.
For a drug insert, ask multiple follow-up questions, such as "What are the main indications for this drug?" followed by "What are the contraindications?", observing the fluency and coherence of the conversation.
Use queries that include specific numerical ranges (e.g., "find all cases with lesion area greater than 10 cm²") to verify whether the system can correctly extract and compare numerical information from documents.
Test uploading a document containing many specialized terms and abbreviations (e.g., "IL-17A," "BSA") to check the conversation system's understanding and retrieval accuracy for these terms.

The values provided are common starting points. Measure them 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.