Proof of Concept

A simple proof of concept (POC) was developed to demonstrate how an automation and AI-based solution can effectively integrate into real clinical environments. This prototype allows users to quickly and contextually check a patient’s recorded medical visits (including relevant data such as weight, height, consultation date, and clinical notes) and verify whether a healthcare professional is available for a new appointment. The solution, built using visual workflows in n8n and connected to a structured medical database, shows how accurate and helpful responses can be delivered through a channel like Telegram (without the need for complex apps or multiple steps). All of this was achieved by combining tools like Pinecone for semantic search and Azure OpenAI for natural language understanding, resulting in a smooth, user-centered conversational experience.

Creation Embeddings Flow

For the AI assistant to provide useful and contextual responses, all information related to medical appointments and clinical records must be transformed into a format that allows it to truly understand the content (not just read it literally). That’s why the first automation focuses on converting this data into embeddings (numerical representations that capture the meaning of the text).

This process runs automatically every hour, ensuring that any new data (such as a recent appointment or updated clinical note) is quickly processed and indexed. The workflow begins with an API call that retrieves the most recent clinical records or, if it’s the first run, the entire medical history. The data then goes through a processing and cleanup stage before being sent to Azure OpenAI, where the corresponding embeddings are generated.

These vectors are stored in Pinecone, a system specialized in semantic search, allowing the AI assistant to retrieve relevant information accurately (even when the user doesn’t phrase the question exactly as it was recorded).

Thanks to this preparation step, the assistant can respond with specific details about diagnoses, consultation dates, or previously recorded information (all without the user having to manually search through their history). This approach not only improves the user experience, it also ensures that the information is available at the right time and communicated in a natural way.

Figure 1: Creation Embeddings Flow

Once the SQL query is executed and the clinical data is retrieved (including patient name, appointment date, medical notes, professional details, and vital signs), the records go through a transformation process to prepare them for embedding generation. This step converts the content of each appointment into a numerical format (one that the language model can understand), enabling more accurate contextual searches and more relevant responses when the assistant queries the information.

Figure 2: Clinical Data Embedding Pipeline, SQL Section

The vectorization process was carried out using the Pinecone Vector Store node, which handles the storage of the generated embeddings in a database specifically designed for high-speed semantic searches. This step ensures that the clinical information is organized in a format the assistant can easily query (even when the user’s questions don’t exactly match the original wording), significantly improving the accuracy and usefulness of each response.

Figure 3: Clinical Data Embedding Pipeline, Creation of the Chunks