Data Characteristics in this Domain
Pharmacovigilance data in telemedicine primarily originates from patient records uploaded via online consultations, health monitoring devices, smart wearables, and healthcare provider entries in electronic medical record systems. Data updates are frequent, typically real-time after each patient consultation, medication record, or device data upload. Document structures are diverse, including unstructured patient self-reported text, structured medication record forms, semi-structured medical report PDFs, and biometric data streams from wearables. In addition to common drug information like drugName, dosage, and frequency, telemedicine-specific fields such as teleconsultationId, deviceSensorData, and patientReportedOutcome are present. Units include milligrams, milliliters, times/day, bpm, and mmHg.
Constraints Imposed by Data Characteristics on Model Integration and Configuration
High-frequency updates of telemedicine data require the knowledge base to support efficient incremental synchronization. This prevents the model from making decisions based on outdated information. Diverse and heterogeneous data structures necessitate more complex cleaning and standardization processes during data preprocessing. Examples include entity extraction from patient self-reported text and structured parsing of PDF reports. The high proportion of unstructured text demands advanced model capabilities for context and semantic understanding, requiring more refined text segmentation strategies. Device sensor data, often time-series, needs specialized processing modules for feature extraction, with results integrated into the knowledge base. Telemedicine-specific fields, such as teleconsultationId, require the model to accurately identify and use them to link data from different sources, ensuring comprehensive pharmacovigilance analysis.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 2048 | Patient descriptions in telemedicine can be lengthy; sufficient context is needed for semantic understanding. |
Chunk size (Segment Length) | 400–600 characters (characters) | Balances semantic completeness and recall efficiency, preventing dilution of information in long paragraphs. |
Recall count (Recall Count) | Top 8 entries (top 8) | Teleconsultations are complex; more relevant information aids in identifying adverse reactions. |
Similarity threshold (Similarity Threshold) | 0.75 | Ensures accuracy of recalled content and reduces interference from irrelevant information. |
Rerank result count (Reranked Return Count) | Top 3 entries (top 3) | After reranking, selecting the few most relevant items improves model response quality. |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds (seconds) | Processing large or complex medical report PDFs can be time-consuming. |
Three Common Pitfalls
- The model fails to proactively query the knowledge base after receiving a message, providing a generic answer instead. This occurs when the
systemPromptdoes not explicitly instruct the model to query the knowledge base under specific conditions. - The voice input function returns an
Unsupported audio formaterror. This likely indicates an incompatibility between the audio encoding format uploaded by the client and the formats supported by the server, such as a non-standard sampling rate or bit rate. - Knowledge base query results do not match expectations, failing to recall relevant adverse reaction information. This often results from an unreasonable text segmentation strategy, where critical information is truncated or semantic units are split, affecting vector embedding quality.
Verification of Configuration
- Submit queries containing typical drug adverse reaction descriptions. Observe if the model accurately cites relevant drug instructions or pharmacovigilance information from the knowledge base.
- Upload a PDF report with complex medication records and patient symptoms. Check if the knowledge base can correctly parse and extract key fields, such as
adverseEventordrugInteraction. - Simulate a telemedicine consultation process. Input patient-reported symptoms and verify if the model, after speech-to-text conversion, triggers the correct knowledge base query based on the text content.
- In a test environment, call the
get_knowledge_base_detailsAPI to check the knowledge base index status, ensuring all uploaded data has been successfully indexed.
Note: The values provided are common starting points. Measure against your own samples to determine optimal settings.
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.