Model Integration and Configuration for Pharmacovigilance

Data Characteristics in Pharmacovigilance

Pharmacovigilance data primarily originates from post-market surveillance reports, clinical trial data, medical literature, social media, and patient self-reports. This data often exists as unstructured text, such as patient medical records, adverse drug reaction (ADR) reports, drug inserts, research papers, and regulatory guidelines. Data updates frequently, especially during the early stages of a new drug's release or when new safety signals emerge. Document structures are complex and may contain medical terminology, abbreviations, dosage units (e.g., mg, ml, IU), time units (e.g., days, weeks, months), and condition descriptions. Fields are diverse, covering patient demographics, drug information (batch number, manufacturer), adverse event descriptions, diagnostic results, and medication history, often with ambiguity and inconsistent expressions.

Constraints from Data Characteristics on Model Integration and Configuration

The complexity of pharmacovigilance data imposes specific requirements on model integration and configuration. High update frequency necessitates that the knowledge base supports efficient incremental updates and version management, ensuring the model always reasons based on the latest information. The prevalence of unstructured text and medical terminology requires the model to possess strong text understanding and entity recognition capabilities, particularly when handling abbreviations and synonyms. For example, PT can refer to "patient" or "prothrombin time," requiring the model to have sufficient contextual awareness. Diverse fields and complex document structures demand more refined text segmentation strategies during data preprocessing to prevent critical information from being truncated or semantic loss. Additionally, accurate identification and standardization of units like dosage and time are crucial for the model to avoid misleading information when generating summaries or answers.

Configuration Guidelines

Configuration Item	Suggested Value	Rationale
`Chunk size` (Segment Length)	`500-800 characters`	Pharmacovigilance reports often contain detailed descriptions; this length helps preserve contextual integrity and reduces semantic fragmentation.
`Chunk overlap` (Segment Overlap)	`100-150 characters`	Ensures critical information at segment boundaries is not lost due to splitting, improving retrieval recall.
`maxContext`	`3000-4000 tokens`	Pharmacovigilance queries typically require a longer context to understand complex conditions and medication history, preventing information truncation.
`Recall count` (Recall Count)	`8-12 items`	Increasing the recall count helps cover more potentially relevant adverse event reports or medical literature.
`Similarity threshold` (Similarity Threshold)	`0.75-0.85`	The medical field demands high accuracy; increasing the threshold helps filter out irrelevant or weakly relevant document segments.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Processing large PDF medical literature or reports requires a longer parsing time to avoid timeouts.

Common Configuration Errors

Empty Model Stream Response: This typically results from incorrect model service configuration, such as an invalid API Key or incorrect Endpoint address, preventing FastGPT from establishing an effective connection with the model and retrieving a response.
Image Parsing Failure: When using multimodal models to process images, if a prompt indicates an inability to download an image, the FastGPT runtime environment's network configuration may restrict access to external image links, or the image URL may have an anti-hotlinking mechanism.
Vector Model Normalization Mismatch: When integrating a new vector model (e.g., Doubao's embedding), if its output vectors are not normalized and FastGPT does not have vector normalization enabled, it leads to abnormal vector similarity calculation results, affecting retrieval effectiveness.

Verification Steps

Upload a typical adverse event report PDF file. Observe if file parsing is successful and if segmentation results are reasonable. Check if critical information (e.g., drug name, adverse reaction, dosage) is completely retained.
Ask specific questions about drug adverse reactions, for example, "What are the common gastrointestinal adverse reactions of aspirin?" Check if the model's answer accurately cites relevant information from the knowledge base and verify the sources.
Simulate the model's load during peak periods. Check if timeout configurations like PARSE_FILE_TIMEOUT_SECONDS effectively handle large files or complex queries, preventing service interruptions due to timeouts.
Verify if the model can correctly identify and provide relevant results when processing queries containing medical abbreviations and synonyms. For example, query "Precautions for hypertensive patients taking ACEI drugs" and observe if the model can identify ACEI and link it to relevant drugs.

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.