Data Characteristics in This Category
Surgical robot pharmacovigilance data primarily originates from Post-Market Surveillance (PMS) reports, clinical trial reports, Real-World Evidence (RWE), and medical device adverse event databases. This data typically exists in a mixed format, combining structured data (e.g., database records, tables) and unstructured data (e.g., free-text descriptions, imaging reports). Adverse event reports are usually real-time or near real-time, while PMS reports may be compiled quarterly or annually. Document structures include medical device registration certificates, product manuals, user manuals, maintenance logs, and adverse event reporting forms (e.g., FDA's MedWatch 3500A or EU's EUDAMED reports). Fields and units involve device model, batch number, serial number, usage date, patient demographic information, adverse event description, event occurrence time, device operator, device fault codes, and maintenance records. Time fields are precise to the hour, dosage units are typically international units or milligrams, and device parameters like power and pressure have defined engineering units.
Constraints on "Multi-Turn Conversations and Prompts" Imposed by These Characteristics
The high real-time nature and mixed structure of surgical robot pharmacovigilance data impose specific requirements on the accuracy of multi-turn conversations and the construction of prompts. Real-time reporting necessitates that the conversation system can quickly ingest and process new information. Prompt design must consider timeliness to avoid referencing outdated data. The mixed data structure means prompts must extract precise values from structured fields and understand detailed adverse event descriptions in unstructured text. For example, prompts need to identify key symptoms like "device jamming" or "accidental touch" from free text. Furthermore, the extensive use of specialized terminology and acronyms (e.g., "AE" for Adverse Event, "MAUDE" for FDA's Manufacturer and User Facility Device Experience database) requires prompts to effectively utilize glossaries and ontological knowledge to reduce ambiguity. Conversations need to guide users to provide critical identification information, such as device serial numbers and batch numbers, to ensure query specificity and avoid generalized responses.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
maxContext | 8 | Ensures the model effectively correlates descriptions of adverse event details from previous turns, maintaining conversational coherence. |
Chunk size | 512 characters | Balances semantic completeness of text with recall efficiency, suitable for report texts containing detailed descriptions. |
Recall count | 7 entries | Considering that adverse event reports may involve multiple related factors, increasing the number of recall items helps ensure comprehensiveness. |
Similarity threshold | 0.75 | For specialized terminology and specific event descriptions, a higher threshold ensures the precision of recalled content. |
Rerank result count | 3 entries | When a large amount of content is recalled, re-ranking to highlight the 3 most relevant items improves efficiency for users to obtain key information. |
SYSTEM_PROMPT | Calibrate based on actual measurements | Must include a clear role setting, such as "As a surgical robot pharmacovigilance assistant, focus on providing accurate adverse event information and device-related consultations." |
Three Common Pitfalls
- Symptom: The model frequently reports "connection error" or "service unavailable" during conversations. Reason: Network configuration issues with the model service interface, such as firewall restrictions or improper proxy settings, prevent FastGPT from properly calling external model APIs.
- Symptom: The model cannot accurately extract device serial numbers or batch information from adverse event reports. Reason: The prompt does not explicitly guide the model to identify and extract key identifiers with specific formats (e.g., alphanumeric combinations), or the knowledge base lacks training data for relevant entity recognition.
- Symptom: When a user asks about image content, the model only returns a text description and cannot display the image. Reason: Images in the knowledge base are not properly indexed or stored as a URL that the model can reference, and the prompt does not include instructions to return image links or identifiers.
How to Confirm Proper Configuration
- Conduct simulated adverse event report queries. Ensure the model accurately identifies and extracts key fields such as device model, event type, and occurrence time from unstructured text.
- Test multi-turn conversation scenarios. For example, first ask about "device jamming," then follow up with "which batch was involved." Verify the model can guide the user to provide necessary information while maintaining context coherence.
- Query the adverse event database using different keyword combinations (including specialized terminology and colloquial descriptions). Check the relevance and accuracy of recall results. Ensure the similarity threshold and the number of recall items are appropriately set.
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.