Skip to main content
Glama
FHIR MCP Server

FHIR MCP Server

by the-momentum

Contact us Visit Momentum MIT License

πŸ“‹ Table of Contents

Related MCP server: FHIR MCP Server

πŸ” About The Project

FHIR MCP Server implements a complete Model Context Protocol (MCP) server, designed to facilitate seamless interaction between LLM-based agents and a FHIR-compliant backend. It provides a standardized interface that enables full CRUD operations on FHIR resources through a comprehensive suite of tools - accessible from MCP-compatible clients such as Claude Desktop, allowing users to query and manipulate clinical data using natural-language prompts.

✨ Key Features

  • πŸš€ FastMCP Framework: Built on FastMCP for high-performance MCP server capabilities

  • πŸ₯ FHIR Resource Management: Full CRUD operations for all major FHIR resources

  • πŸ“„ Intelligent Document Processing: AI-powered document ingestion and chunking for multiple formats including TXT, CSV, JSON, and PDF

  • πŸ” Semantic Search: Advanced document search using vector embeddings (via Pinecone)

  • 🧠 RAG-Ready: Retrieval-Augmented Generation pipeline with context-aware document queries

  • πŸ” Secure Authentication: OAuth2 token management for FHIR API integration

  • πŸ“Š LOINC Integration: Standardized medical terminology lookup and validation

  • 🐳 Container Ready: Docker support for easy deployment and scaling

  • πŸ”§ Configurable: Extensive .env-based configuration options

πŸ—οΈ Architecture

The server is built with a modular architecture:

  • MCP Tools: Dedicated tools for selected FHIR resource types, with others handled by a generic tool

  • Fhir Server Client: Handles FHIR API communication and authentication (OAuth2 and more planned)

  • RAG Services: Embedding-based document processing and semantic retrieval

  • Vector Store: Pinecone integration for similarity-based search

  • LOINC Client: Integration with LOINC API for terminology resolution and validation

πŸ’‘ Demo

This demo shows how Claude uses the fhir-mcp-server to communicate with a FHIR server (in this case Medplum) to answer questions. You will see, among other things:

  • utilization of the request_patient_resource tool which retrieves basic patient information

  • utilization of the request_condition_resource tool to answer the question whether any of the previously diagnosed diseases may cause symptoms that the patient is currently complaining about

  • utilization of the request_medication_resource, request_encounter_resource, request_generic_resource tools to answer the question whether the patient has already received any treatment for hypertension

You can observe how Claude automatically selects the tools worth using to answer the question based on the user's query.

https://github.com/user-attachments/assets/3a3a8ed3-f881-447d-af03-5f24432a2cdd

Here you can observe how Claude first uses the tool searching for LOINC codes for the lipid panel specific codes, but not finding any related observations in FHIR server, it repeats the search for individual biomarkers that make up such a panel.

https://github.com/user-attachments/assets/2fb39801-d5d6-4461-bedd-9f58ab4d52ec

Developers working with FHIR often need to generate specific test data to validate FHIR server functionality, such as search capabilities and data relationships. While you can use Synthea to generate synthetic data and then manually import the resulting bundles to your server, fhir-mcp-server streamlines this process by allowing you to generate and deploy test data directly through Claude.

This eliminates the typical workflow of running synthea separately, downloading bundles, and manually importing them to your FHIR server. Instead, you can create targeted test scenarios, generate appropriate synthetic data, and populate your server all within Claude's interface.

https://github.com/user-attachments/assets/d87da1d8-6401-4a9e-a6f0-50ba23396e12

Note: fhir-mcp-server was not designed with this use case in mind, so as you'll see in the demo, it doesn't work perfectly - what can be observed, however, is how well the LLM handles using trial and error to correct any wrong choices.

πŸš€ Getting Started

Follow these steps to set up FHIR MCP Server in your environment.

Prerequisites

  • Docker (recommended) or uv: For dependency management

    πŸ‘‰ uv Installation Guide

  • FHIR Server Account: Access to FHIR API (e.g. Medplum)

  • Pinecone API key (required for document search): Enables vector-based search over processed documents. Without it, semantic retrieval features will be unavailable.

    πŸ‘‰ Create Pinecone Account

  • LOINC Account (optional): Enables retrieval of the latest LOINC codes from the official API. Without it, the system relies on static or language model-inferred codes, which may be outdated or imprecise.

    πŸ‘‰Create LOINC Account

Installation & Setup

  1. Clone the repository:

    git clone https://github.com/the-momentum/fhir-mcp-server
    cd fhir-mcp-server
  2. Set up environment variables:

    cp config/.env.example config/.env

    Edit the config/.env file with your credentials and configuration. See Environment Variables

  3. Install Dependencies

    For Docker-based execution run:

    make build

    For uv-based execution run:

    make uv
  4. Update the MCP Client configuration

    e.g. Claude Desktop -> edit claude_desktop_config.json

  • Docker

    {
       "mcpServers": {
          "docker-mcp-server": {
             "command": "docker",
             "args": [
                "run",
                "-i",
                "--rm",
                "--init",
                "--name",
                "fhir-mcp-server",
                "--mount", // optional - volume for reload
                "type=bind,source=<your-project-path>/app,target=/root_project/app", // optional - volume for reload
                "--mount",
                "type=bind,source=<your-project-path>/config/.env,target=/root_project/config/.env",
                "-e", "TRANSPORT_MODE=stdio", // Set transport mode: stdio, http, or https
                "mcp-server:latest"
             ]
          }
       }
    }

    Make sure to replace <your-project-path> with the actual path to your installation

  • uv

    Firstly, get uv path from terminal:

    • Windows:

      (Get-Command uv).Path
    • MacOS/Linux:

      which uv

    Then, update config file:

    {
       "mcpServers": {
          "uv-mcp-server": {
             "command": "uv",
             "args": [
                "run",
                "--frozen",
                "--directory",
                "<your-project-path>",
                "start"
             ],
             "env": {
             "PATH": "<uv-bin-folder-path>"
             }
          }
       }
    }

    Make sure to replace with the actual uv path (to bin folder)

  1. Restart MCP Client

    After completing all of the above steps, restart the MCP Client to apply the changes. In some cases, you may need to terminate all related processes using Task Manager or your system's process manager. This ensures that:

    • The updated configuration is properly loaded

    • Environment variables are correctly applied

    • The FHIR MCP client initializes with the correct settings

πŸ”§ Configuration

πŸ” Security & Encryption

The FHIR MCP Server includes built-in encryption infrastructure to protect sensitive configuration values. Sensitive fields like API keys and passwords are automatically encrypted and decrypted at runtime.

You are allowed to store passwords as a plain text, but if you want to have them encrypted, follow the instruction below.

Setting Up Encryption

For most users, use the automated setup script:

# uv method
uv run scripts/cryptography/setup_encryption.py

# docker method
docker exec fhir-mcp-server uv run scripts/cryptography/setup_encryption.py

This script will:

  1. Check for MASTER_KEY in config/.env and generate one if needed

  2. Automatically encrypt all sensitive values (LOINC_PASSWORD, FHIR_SERVER_CLIENT_SECRET, PINECONE_API_KEY)

  3. Update your .env file with encrypted values

  4. Skip empty variables and already encrypted values

  1. Generate a Master Key:

    # uv method
    uv run scripts/cryptography/generate_master_key.py
    
    # docker method
    docker exec fhir-mcp-server uv run scripts/cryptography/generate_master_key.py

    Put that key as a MASTER_KEY environment variable in .env.

  2. Encrypt Sensitive Values:

    # uv method
    uv run scripts/cryptography/encrypt_setting.py "your_secret_value"
    
     # docker method
    docker exec fhir-mcp-server uv run scripts/cryptography/encrypt_setting.py "your_secret_value"
  3. Decrypt Values (for verification):

    # uv method
    uv run scripts/cryptography/decrypt_setting.py "encrypted_value"
    
     # docker method
    docker exec fhir-mcp-server uv run scripts/cryptography/decrypt_setting.py "encrypted_value"

Encrypted Configuration Fields

The following fields are automatically encrypted when using EncryptedField:

  • FHIR_SERVER_CLIENT_SECRET - OAuth2 client secret for FHIR server

  • LOINC_PASSWORD - LOINC account password

  • PINECONE_API_KEY - Pinecone API key for vector search

Environment Variables

Variable

Description

Example Value

Encryption

MASTER_KEY

Master encryption key

gAAAAABl...

Required

FHIR_SERVER_HOST

FHIR API host URL

https://api.medplum.com

No

FHIR_BASE_URL

FHIR base path

/fhir/R4

No

FHIR_SERVER_CLIENT_ID

OAuth2 client ID for FHIR

019720e7...

No

FHIR_SERVER_CLIENT_SECRET

OAuth2 client secret for FHIR

gAAAAABl...

Yes

LOINC_ENDPOINT

LOINC API search endpoint

https://loinc.regenstrief.org/searchapi/loincs

No

LOINC_USERNAME

LOINC account username

loinc-user

No

LOINC_PASSWORD

LOINC account password

gAAAAABl...

Yes

PINECONE_API_KEY

Pinecone API key

gAAAAABl...

Yes

EMBEDDING_MODEL

Hugging Face embedding model name

NeuML/pubmedbert-base-embeddings

No

πŸ› οΈ MCP Tools

The FHIR MCP Server provides a comprehensive set of tools for interacting with FHIR resources and document management:

FHIR Resource Tools

Tool

Resource Type

Description

request_patient_resource

Patient

Manage patient demographic and administrative information

request_observation_resource

Observation

Handle clinical measurements and assessments

request_condition_resource

Condition

Manage patient problems and diagnoses

request_medication_resource

Medication

Handle medication information and orders

request_immunization_resource

Immunization

Manage vaccination records

request_encounter_resource

Encounter

Handle patient visits and interactions

request_allergy_intolerance_resource

AllergyIntolerance

Manage patient allergy information

request_family_member_history_resource

FamilyMemberHistory

Handle family health history

request_generic_resource

Any FHIR Resource

Operate on any FHIR resource not covered by specific tools

Document Management Tools

Tool

Description

request_document_reference_resource

Manage FHIR DocumentReference resources

add_document_to_pinecone

Ingests documents into the vector database for semantic search

search_pinecone

Performs semantic search across indexed documents using vector embeddings

LOINC Terminology Tools

Tool

Description

get_loinc_codes

Retrieves standardized LOINC codes for medical observations and laboratory tests

Tool Features

  • Full Resource Management: All FHIR resource tools support Create, Read, Update, and Delete operations

  • Data Validation: Tools enforce FHIR resource validation and prevent data corruption

  • Error Handling: Comprehensive error responses with detailed failure information

  • Security: OAuth2 authentication and proper access control for all operations

  • Semantic Search: AI-powered document search using vector embeddings

  • Multi-format Support: Document ingestion supports TXT, PDF, CSV, and JSON formats

πŸ—ΊοΈ Roadmap

We're continuously enhancing FHIR MCP Server with new capabilities. Here's what's on the horizon:

  • Extended Authentication Options: In addition to OAuth2 (already supported), we plan to add support for other authentication methods for connecting to FHIR servers

  • Expanded File Format Support for RAG: Extend document ingestion capabilities to support additional formats

  • Table-Aware Document Chunking: Improve the document chunking pipeline by detecting tables in documents and treating them as separate, atomic chunks.

  • OCR Support for Scanned Documents: Implement Optical Character Recognition capabilities to enable extraction of text from scanned PDFs and image files before chunking and indexing

Have a suggestion? We'd love to hear from you! Contact us or contribute directly.

πŸ‘₯ Contributors

πŸ“„ License

Distributed under the MIT License. See MIT License for more information.


Available Tools

13 tools
add_document_to_pineconeA

IMPORTANT: Always inform the user at the beginning of your response that this search operation may take some time because the embedding model will be loaded into cache. Adds a document to the Pinecone vector index for the specified FHIR DocumentReference ID.

This tool should be used to ingest new documents into the Pinecone index.

Rules: - If you cannot determine the format of the document fitting the format from the list, provide the format as None. - After adding the document, the Pinecone index may take up to 1 minute to update before the document is searchable.

Args: url (str): The URL of the document to be added. fhir_document_id (str): The ID of the FHIR DocumentReference resource corresponding to the document. format (literal | None): The format of the document.

Returns: str: Confirmation message that the document was added or already exists. PineconeError: Error object with a message if the operation fails.

Inputs: document (required). Returns: result

get_loinc_codesA

Get the most relevant LOINC codes for a given observation name.

The function automatically:

  • Filters for STATUS="ACTIVE" codes only

  • Sorts by COMMON_TEST_RANK (lower rank = more commonly used)

  • Returns codes in popularity order (most common first)

Your job is to:

  1. Analyze returned codes for semantic relevance to the search query

  2. Balance clinical popularity with semantic matching

  3. Select codes that best match the intended observation

Strategy:

  1. Start with default parameters (max_codes=5, max_fetch=50)

  2. Check if result contains "Error" key in first element

  3. If "Authentication failed" or "Authorization" error:

    • STOP using this tool immediately.

    • Do not retry with different parameters.

    • Report the authentication error to the user.

    • Suggest they check their LOINC API credentials.

    • Ask the user if they want to use your knowledge to find a LOINC code and wait for the confirmation. Add warning that this may cause wrong results.

  4. If "No active LOINC codes found in current fetch":

    • Increase max_fetch progressively (50β†’100β†’200β†’RecordsFound).

    • Keep trying until max_fetch >= RecordsFound or you find active codes.

  5. If "No LOINC codes found": Try alternative search terms or report failure.

  6. If you get codes but they don't semantically match your query:

    • Increase max_codes to see more options.

    • Look for better matches in COMPONENT, SHORTNAME, LONG_COMMON_NAME fields.

Rules:

  • Function returns codes sorted by popularity - YOU decide which are most relevant.

  • Don't automatically pick the first (most common) codes.

  • Prioritize semantic relevance: exact matches in COMPONENT > SHORTNAME > partial matches.

  • Balance popularity with relevance (very rare codes might not be clinically useful).

  • Keep increasing max_fetch until you exhaust all available records (max_fetch >= RecordsFound).

  • Increase max_codes only when you need more options to find better semantic matches.

Args: component_name: The name of the observation to get the LOINC code for (i.e. "glucose"). max_codes: The maximum number of LOINC codes to return. max_fetch: The maximum number of LOINC codes to fetch from the API. Returns: LOINC codes sorted by popularity - you must select the most semantically relevant ones.

Inputs: component_name (required), max_codes, max_fetch. Returns: result

request_allergy_intolerance_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR AllergyIntolerance resource. Rules: - When creating or updating an allergy intolerance, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting an allergy intolerance, ask the user for confirmation with details of the allergy intolerance and wait for the user's confirmation. - Provide links to the app (not api) allergy intolerance resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/AllergyIntolerance", "/AllergyIntolerance?patient=Patient/123") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_condition_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Condition resource. Rules: - When creating or updating a condition, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a condition, ask the user for confirmation with details of the condition and wait for the user's confirmation. - Provide links to the app (not api) condition resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Condition", "/Condition?patient=Patient/123") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_document_reference_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR DocumentReference resource. Rules: - When creating or updating a document reference, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a document reference, ask the user for confirmation with details of the document and wait for the user's confirmation. - Provide links to the app (not api) document reference resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/DocumentReference", "/DocumentReference?patient=Patient/123") body: Optional JSON data for POST/PUT requests

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_encounter_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Encounter resource. Rules: - When creating or updating an encounter, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting an encounter, ask the user for confirmation with details of the encounter and wait for the user's confirmation. - Provide links to the app (not api) encounter resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Encounter", "/Encounter?patient=Patient/123") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_family_member_history_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR FamilyMemberHistory resource. Rules: - When creating or updating a family member history, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a family member history, ask the user for confirmation with details of the family member history and wait for the user's confirmation. - Provide links to the app (not api) family member history resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/FamilyMemberHistory", "/FamilyMemberHistory?patient=Patient/123") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_generic_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations on any FHIR resource ONLY if the other tools are not applicable.

Rules: - When creating or updating a resource, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a resource, ask the user for confirmation with details of the resource and wait for the user's confirmation. - Provide links to the app (not api) resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_immunization_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Immunization resource. Rules: - When creating or updating an immunization, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting an immunization, ask the user for confirmation with details of the immunization and wait for the user's confirmation. - Provide links to the app (not api) immunization resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Immunization", "/Immunization?patient=Patient/123") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_medication_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Medication resource. Rules: - When creating or updating a medication, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a medication, ask the user for confirmation with details of the medication and wait for the user's confirmation. - Provide links to the app (not api) medication resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Medication", "/Medication?code=aspirin") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_observation_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Observation resource.

IMPORTANT: Before fetching observations that require LOINC codes:

  1. First use get_loinc_code() tool to find appropriate LOINC codes

  2. Then use this tool to fetch the observation with the LOINC code

Rules: - When creating or updating an observation, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting an observation, ask the user for confirmation with details of the observation and wait for the user's confirmation. - Provide links to the app (not api) observation resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Observation?subject:Patient.name=Homer%20Simpson") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

request_patient_resourceA

Makes an HTTP request to the FHIR server. Use this tool to perform CRUD operations only on the FHIR Patient resource. Rules: - When creating or updating a patient, use only the data explicitly provided by the user. - Do not guess, auto-fill, or assume any missing data. - When deleting a patient, ask the user for confirmation with details of the patient and wait for the user's confirmation. - Provide links to the app (not api) patient resource in the final response.

Args: method: HTTP method (GET, POST, PUT, DELETE) path: Resource path (e.g., "/Patient", "/Patient?name=John%20Doe") body: Optional JSON data for POST/PUT requests)

Returns: JSON response from the FHIR server

Inputs: request (required). Returns: result

search_pineconeA

IMPORTANT: Always inform the user at the beginning of your response that this search operation may take some time because the embedding model will be loaded into cache. Searches the Pinecone vector index for information related to the given document by FHIR DocumentReference ID.

Use this tool when the user requests information from the documents, notes, etc.

Rules: - Firstly, prepare fhir_document_id by running the appropriate tool. - If the error message "Document does not exist in Pinecone index" is returned, automatically trigger the 'add_document_to_pinecone' tool to add the missing document to the index. - Translate the user's query into the language of the document before performing the search. - Base all answers strictly on the content found in the Pinecone index documents. - If the user's question is unrelated to the indexed documents, respond that the information is not available in the documents. - If the query is unclear or ambiguous, ask the user to clarify or provide more details. - You can modify the query to make it more specific and relevant to the document.

Args: query (str): The user's search query. fhir_document_id (str): The ID of the FHIR DocumentReference resource to search within - it is the same as the FHIR ID of the document. top_k (int, optional): The maximum number of search results to return. Defaults to 10.

Returns: list[PineconeSearchResponse]: List of search results matching the query. PineconeError: Error object with a message if the search fails.

Inputs: query (required), fhir_document_id (required), top_k. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 13 tools

    Disambiguation4/5

    The request_*_resource tools are clearly separated by FHIR resource type, and request_generic_resource is explicitly a fallback for resources without a dedicated tool. However, the similar structure and the note in request_observation_resource referencing a non-existent 'get_loinc_code' tool could cause slight confusion.

    Naming Consistency4/5

    The CRUD tools follow a strong request_<resource>_resource pattern, making the set predictable. The utility tools (get_loinc_codes, add_document_to_pinecone, search_pinecone) diverge in style but are still verb-first and understandable.

    Tool Count5/5

    13 tools is well within the ideal range for a FHIR server covering multiple resource types. The number is neither bloated nor thin, given the broad scope of FHIR.

    Completeness4/5

    The set covers CRUD for the most common FHIR resources plus a generic fallback for all others, and adds LOINC lookup and document search. Minor gaps like a dedicated search/query tool or terminology service exist, but the generic resource tool covers most needs.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables LLM-based agents and MCP-compatible clients to perform full CRUD operations on FHIR resources and query or manipulate clinical data using natural-language prompts. It also supports document ingestion, semantic search, LOINC terminology lookup, and OAuth2 authentication for FHIR APIs.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables LLM-based agents and MCP-compatible clients to perform full CRUD operations on FHIR resources and query or manipulate clinical data using natural-language prompts. It also supports document ingestion, semantic search, LOINC terminology lookup, and secure FHIR API authentication.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables LLM-based agents to connect to a FHIR-compliant healthcare backend and query or manipulate clinical data with natural-language prompts. Provides full CRUD operations on FHIR resources, AI-powered document ingestion and semantic search, RAG-ready retrieval, and LOINC terminology lookup.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables LLM-based agents and MCP-compatible clients to query and manipulate clinical data on a FHIR-compliant backend using natural-language prompts, with full CRUD operations, document ingestion, semantic search, and LOINC terminology support.
      -