Data Characteristics in This Category
Contract Sales Organizations (CSOs) in pharmacovigilance primarily use data from their pharmaceutical partners. This includes clinical trial reports, real-world studies, post-market adverse event (AE/SAE) reports, literature reviews, and patient feedback. This data often exists as unstructured text, such as medical reports, patient diaries, and transcribed phone calls. It also includes structured data like coded information (MedDRA, WHO-ART) from adverse event databases. Data updates frequently, especially early in a drug's lifecycle, with new adverse event reports arriving daily or hourly. Document structures vary, from standardized CIOMS I forms to free-form narrative reports. Fields and units are specialized. For example, adverse event incidence rates are often expressed as "per thousand patient-years" or "per million doses." Drug dosages use various units, requiring handling of milligrams, micrograms, and other units.
Constraints from "Tool Calling and Plugins" Due to These Characteristics
The high timeliness of CSO pharmacovigilance data requires tool calling plugins to process data in real-time or near real-time. This ensures timely identification and reporting of adverse events. Diverse unstructured data sources and document formats necessitate robust text parsing capabilities in plugins. Plugins must accurately extract key information from free text, such as drug names, adverse events, patient characteristics, and occurrence times. This directly impacts the complexity and robustness of information extraction plugins. Specialized fields and units, like MedDRA codes and drug dosage units, require plugins to precisely identify and map to predefined ontologies or dictionaries during data standardization and conversion. This prevents data errors caused by unit confusion. High update frequency means knowledge bases require frequent incremental updates or re-indexing. Tool calling plugins must trigger or adapt to this dynamic update mechanism, ensuring analysis always uses the latest data.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxTokens | 2048 | Accommodates lengthy medical reports, ensuring text completeness and preventing truncation of critical information. |
temperature | 0.3 | Pharmacovigilance emphasizes factual accuracy and consistency. A lower temperature reduces model hallucination. |
recallTopK | 10 | Recalls a sufficient number of relevant adverse event reports from the knowledge base, improving event correlation. |
similarityThreshold | 0.75 | A high threshold ensures recalled knowledge snippets are highly relevant to the query, filtering noise. |
parserConfig.extractKeywords | True | Automatically extracts key medical terms, drug names, and adverse events, aiding subsequent analysis. |
pluginTimeoutSeconds | 60 seconds | Allows ample execution time, considering potential delays when processing large text data and external API calls. |
Common Pitfalls
- When calling external APIs, the log shows "model stream response is empty." This typically occurs when the external service returns an unexpected empty response or data in an incorrect format, preventing FastGPT from parsing it.
- The model configured in the knowledge base reference differs from the model used during actual calls. This leads to reduced generation quality or logical errors, often because application configuration priorities or inheritance relationships are misunderstood.
- Plugins make errors when processing medical terms and dosage units. This results in extracted information not matching the original report. This usually happens because the plugin is insufficiently trained or lacks integration with specialized medical dictionaries and unit conversion rules.
Verification Steps
- Check the execution status of each tool calling plugin in the system logs. Ensure no abnormal errors or timeouts occur, and verify that the returned data structure matches expectations.
- For specific adverse event reports, manually verify that key information extracted by the plugin (e.g., drug name, adverse event, dosage) precisely matches the original document. Compare this information with MedDRA codes.
- Design a set of test cases containing typical medical terms and complex units. Run the plugin and check its standardized output to ensure accuracy in unit conversion and term mapping.
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.