Data Characteristics
Clinical trial pre-screening data comes from global clinical trial registries (e.g., ClinicalTrials.gov, WHO ICTRP) and trial protocols published by pharmaceutical companies and research institutions. This data often combines structured and semi-structured formats. It includes fields such as trial name, disease area, inclusion/exclusion criteria, research centers, contact information, trial phase, and recruitment status. Data updates frequently, typically weekly or daily, as new trials register and existing trial statuses change. Document structures vary; PDF trial protocol documents are common and contain extensive unstructured text descriptions. Inclusion/exclusion criteria are central. These criteria involve medical terminology, disease diagnosis codes (e.g., ICD-10), biomarker results, and medication history. Units cover dosage (mg), time (weeks, months), and physiological indicators (mmHg, mmol/L).
Constraints on Multi-Turn Conversations and Prompts
Frequent data updates require an efficient, automated knowledge base synchronization mechanism. This prevents the model from providing advice based on outdated information. The mix of semi-structured and unstructured data demands robust document parsing and information extraction capabilities. These capabilities must accurately extract and structure inclusion criteria from PDFs for model comprehension. The complexity of medical terminology and diagnostic codes challenges the model's ability to understand user intent and generate professional responses. Pre-loaded medical dictionaries and ontologies are necessary. In multi-turn conversations, users may progressively provide personal health information. The model must accurately link this information to trial inclusion criteria, gradually narrowing the scope of eligible trials. For example, after a user mentions "hypertension," the model needs to ask about "blood pressure control" or "medication history." Missing this information directly impacts pre-screening accuracy. Prompt design must guide users to provide critical information and ensure the model can map this information to complex clinical trial standards.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 2048 token | Ensures sufficient historical information is retained in multi-turn conversations, covering 5-7 rounds of Q&A, to accurately understand user needs and context. |
Chunk size (Segment Length) | 500 characters (characters) | Inclusion criteria entries in clinical trial documents are often long; this length helps maintain the integrity of individual semantic segments. |
Recall count (Recall Count) | Top 10 entries (top 10) | Given the complexity of clinical trial inclusion criteria, recalling more relevant snippets improves matching accuracy. |
Similarity threshold (Similarity Threshold) | 0.75 | Clinical trial pre-screening requires high matching precision. A lower threshold may introduce irrelevant information, while a higher threshold may miss potential matches. |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds (seconds) | Clinical trial PDF documents can be large and complex, requiring sufficient time for parsing and vectorization. |
Rerank result count (Rerank Return Count) | Top 5 entries (top 5) | After reranking, taking the top 5 most relevant results balances accuracy and model input length. |
Common Pitfalls
- The conversation returns "No relevant trials found" or "Insufficient information to determine." This may occur if relevant trial documents in the knowledge base are not parsed correctly, preventing the model from retrieving valid information.
- After the user inputs symptom information, the model fails to ask for critical diagnostic details or medication history. This typically happens when the prompt lacks guidance for deeper medical information extraction, failing to fully leverage the multi-turn conversation mechanism.
- API call responses differ from platform test results. This may be because the knowledge base configuration was not correctly passed or loaded during the API call, causing the model to respond without knowledge base support.
Verification Steps
- Select 5-10 typical patient cases. Simulate the multi-turn conversation flow. Check if the model can accurately converge to a list of eligible clinical trials or ask appropriate follow-up questions based on progressively provided user information.
- Upload 3-5 clinical trial protocols in different formats (e.g., PDF, Docx). Confirm that the document parser correctly extracts key inclusion/exclusion criteria fields and completes within
PARSE_FILE_TIMEOUT_SECONDS. - Use the FastGPT platform for testing. Compare answers to specific clinical trial queries with and without the knowledge base enabled. Ensure the knowledge base effectively intervenes and enhances the professionalism of the answers.
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.