Multiturn Conversation and Prompting for Stem Cell Therapy

Data Characteristics

Stem cell therapy pharmacovigilance data comes from clinical trial reports, real-world evidence (RWE) data, case reports, and regulatory adverse event databases. This data is often structured (e.g., MedDRA-coded adverse events) and unstructured (e.g., physician notes, patient self-reports). Clinical trial data updates regularly during the trial period. Post-market data continuously increases. Document structures are complex. They include patient demographics, treatment plans, stem cell product batch information, adverse event occurrence time, severity, outcome, and causality assessment. Stem cell therapy data uniquely includes fields like cell source, preparation process, and administration route. Adverse event descriptions often involve specific terminology, such as immunogenicity and graft-versus-host disease (GVHD).

Constraints on Multiturn Conversation and Prompting

The specialized and complex nature of stem cell therapy pharmacovigilance data demands high accuracy for multiturn conversations and prompt design. Unstructured text contains many medical terms and abbreviations. The model requires strong semantic understanding. Dynamic data updates require rapid knowledge base synchronization to avoid outdated information. Key fields like stem cell product batch and cell source are critical for tracing adverse events. These must be accurately identified and extracted during conversations. Life safety is involved. Any adverse event query must ensure information rigor and traceability. This limits the dialogue system's freedom in inference and generalization, focusing more on factual retrieval and specific information summarization. Causality judgment often requires combining multi-dimensional data. Prompt design must guide the model through multi-step reasoning.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8`	Ensures sufficient context to trace patient history, treatment plans, and adverse event details across multiple turns.
`recallTopK`	`10`	Increases the number of recalled items to capture various potential adverse event types and related factors in stem cell therapy, enhancing comprehensiveness.
`similarityThreshold`	`0.75`	Guarantees high relevance between recall results and user queries, reducing interference from inaccurate or irrelevant information, especially for specialized terminology matching.
`rerankTopN`	`3`	Reranks recall results to prioritize the top three most relevant pieces of information, improving query efficiency.
`chunkSize`	`500–800 characters`	Adapts to the typical length of medical texts. Provides enough information without being too long, aiding model comprehension.
`temperature`	`0.3`	Reduces the randomness of model-generated answers. Ensures rigor and accuracy in pharmacovigilance scenarios, avoiding hallucinations.

Common Pitfalls

The "thinking..." prompt repeatedly appears in the conversation. This occurs when the code execution node after the AI dialogue node in the workflow takes too long, causing a global response timeout.
Queries for adverse reactions of specific stem cell product batches return empty or incomplete results. This happens if the knowledge base indexing does not adequately consider the independent retrievability of critical fields like batch numbers, or if the chunking strategy truncates key information.
Users attempting to retrieve historical conversation records find that all user session records are returned. This is due to incorrect customUid parameter passing during API calls or a lack of isolation by user ID in backend storage, leading to mixed historical records.

Validation Steps

Simulate multiple complex multiturn conversations involving stem cell product batches, cell sources, and adverse event descriptions. Verify the system's ability to accurately extract key information and provide compliant responses.
Randomly select documents from the knowledge base related to stem cell therapy adverse reactions. Perform keyword searches and cross-reference recall results with original documents. Ensure similarityThreshold and recallTopK parameter settings are appropriate.
For typical pharmacovigilance queries, check system response times and analyze workflow logs. Confirm that node execution order and time consumption meet expectations, with no unnecessary delays or errors.
Simulate user logins with different customUid values. Query historical conversation records to confirm that each user can only access their own session history.

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.