MedDocs AI

Turn messy medical documents into structured, searchable, actionable patient intelligence.

MedDocs AI is an AI-powered medical document processing platform that automatically extracts, organizes, and understands clinical data from uploaded documents — so healthcare teams can focus on care, not data entry.

---

Architecture

graph TB
    subgraph Frontend ["Frontend — Next.js + React :3000"]
        UI[Dashboard]
        UP[Upload]
        MON[Monitoring]
    end

    subgraph Backend ["Backend — FastAPI :8000"]
        API[REST API]
        MON_API[Monitoring Service :8001]
    end

    subgraph AI["AI Pipeline — LangGraph"]
        P[Planner]
        E[Extractor]
        V[Verifier]
        PE[Persist + Embed]
        Q[RAG Query]
        RG[Report Gen]
        RF[Risk Flag]
    end

    subgraph Workers ["Celery Workers"]
        CW[process_document]
        RT[generate_report]
    end

    subgraph Data ["Data Layer"]
        PG[(PostgreSQL + pgvector)]
        RD[(Redis)]
        FS[(File Storage)]
    end

    subgraph LLM ["External LLM"]
        GROQ[Groq API — Llama 3.1/3.3]
    end

    UI --> API
    UP --> API
    MON --> MON_API
    API --> CW
    CW --> P --> E --> V --> PE
    PE --> PG
    PE --> FS
    E --> GROQ
    V --> GROQ
    P --> GROQ
    Q --> PG
    Q --> GROQ
    RG --> GROQ
    RF --> GROQ
    CW --> RD
    MON_API --> RD
    RD --> MON
    CW -.->|Pub/Sub| RD

    style Frontend fill:#e0f2fe,stroke:#0284c7
    style Backend fill:#f0fdf4,stroke:#16a34a
    style AI fill:#fef3c7,stroke:#d97706
    style Workers fill:#fce7f3,stroke:#db2777
    style Data fill:#f3e8ff,stroke:#9333ea
    style LLM fill:#fff7ed,stroke:#ea580c

Loading

---

What it does

Upload any medical document — lab reports, prescriptions, insurance claims, discharge summaries — and MedDocs AI handles the rest.

Intelligent Document Processing

Drop in a PDF or CSV. Our AI pipeline automatically:

• Classifies the document type (lab report, prescription, claim, etc.)
• Extracts structured data — test results, medication details, claim amounts, diagnoses
• Validates extraction quality with automatic retry on low-confidence results
• Indexes everything for instant natural language search

All of this happens in real-time, with live pipeline monitoring so you always know where things stand.

Structured Data, Ready to Use

Once processed, your documents become:

Extracted Data	What you see
Lab Results	Test names, values, units, reference ranges, abnormal flags, dates
Prescriptions	Drug names, dosages, frequencies, prescribing doctors
Insurance Claims	Procedure codes, amounts, status, dates
Clinical Notes	Full extracted text, chunked and embedded for search

Ask Anything

Query patient records in plain English:

"What were John's last 3 hemoglobin levels?" "Which medications is Sarah currently prescribed?" "Show me all abnormal lab results from this month."

MedDocs AI uses RAG (Retrieval-Augmented Generation) to find relevant document chunks and generate accurate, sourced answers.

AI-Generated Reports

Generate comprehensive clinical reports with one click. The AI planner decides which sections to include based on available patient data:

Section	When Included
Patient Overview	Always — demographics, visit context, summary
Lab Results Summary	When lab data exists — key findings, abnormal values
Medication Summary	When prescriptions exist — current meds, dosages
Risk Flags	When open risk flags exist — clinical advisories
Lab Trends	When multiple lab tests available — directional analysis

Each section is independently editable:

• Click Edit to modify any section's content
• Click Regenerate to have AI rewrite a section with fresh context
• Click + Add Note to insert custom clinical notes
• Click Finalize to lock the report
• Click Download PDF to export a formatted clinical report

View Sample Report →

Reports are generated in the background via Celery workers — the UI polls until the report is ready, then displays the full editable canvas.

Risk Detection

Automatically flags potential concerns:

• Drug interactions between current prescriptions
• Abnormal lab trends that warrant attention
• Claims requiring follow-up or appeal

Flags are surfaced in real-time and can be acknowledged or dismissed with documented reasons.

---

How it works

Upload Document
      ↓
  Classify ──→ What type of document is this?
      ↓
  Extract ──→ Pull structured data from the document
      ↓
  Verify ──→ Is the extraction accurate? Retry if needed.
      ↓
  Persist ──→ Save structured data to the patient record
      ↓
  Embed ──→ Index text for semantic search
      ↓
  Done ──→ Document is searchable, queryable, and report-ready

All steps run asynchronously via background workers, with real-time status updates streamed to the dashboard.

---

Multi-Tenant by Design

Every piece of data is scoped to a tenant. Patient records, documents, extracted data, reports — nothing leaks across organizations. MedDocs AI is built for environments where data isolation isn't optional.

---

Built with

Backend

Technology	Purpose
Python 3.14	Backend runtime
FastAPI	Async REST API framework
SQLAlchemy 2.0	ORM (async + sync engines)
Celery	Distributed background task queue
Redis	Task broker + Pub/Sub event streaming
PostgreSQL 16	Primary database
pgvector	Vector similarity search for RAG (384-dim embeddings)
Alembic	Database migrations

AI / ML

Technology	Purpose
LangGraph	Agentic pipeline with conditional routing and state management
LangChain	LLM orchestration abstractions
Groq (Llama 3.1/3.3)	Fast LLM inference for classification, extraction, verification, RAG, reports
sentence-transformers	Local embedding model loading
BAAI/bge-small-en-v1.5	384-dimensional text embeddings for semantic search

Document Processing

Technology	Purpose
pdfplumber	PDF text and table extraction
pytesseract + pdf2image	OCR fallback for scanned documents
pandas	CSV parsing and tabular data manipulation
fpdf2	PDF report generation and export

Frontend

Technology	Purpose
Next.js 15	React framework with App Router
React 19	UI library
TypeScript 5.7	Type-safe development
Tailwind CSS 3.4	Utility-first styling
react-dropzone	Drag-and-drop file uploads
EventSource (SSE)	Real-time pipeline monitoring

Infrastructure

Technology	Purpose
Docker Compose	Container orchestration (PostgreSQL + Redis)
Uvicorn	ASGI server
sse-starlette	Server-Sent Events streaming
Bash scripts	Automated startup orchestration

Patterns & Architecture

Pattern	Implementation
Multi-tenancy	Tenant-scoped data isolation across all 15 database tables
Agentic AI Pipeline	LangGraph StateGraph with planner → extractor → verifier → persist → embed flow
RAG	pgvector cosine search + structured data augmentation for natural language queries
Event-Driven	Redis Pub/Sub → SSE for real-time pipeline status
Background Processing	Celery async tasks with retry/backoff
Rate Limiting	Custom limiter with exponential backoff for LLM API calls
Graceful Degradation	OCR fallback, verification with auto-retry, fallback classifiers

---

Getting Started

Prerequisites

• Docker Desktop
• Python 3.14+
• Node.js 18+
• Groq API key (free at console.groq.com)

Quick Start

# Clone
git clone https://github.com/its-abhishek/meddoc_ai.git
cd meddoc_ai

# Configure
echo "GROQ_API_KEY=your_key_here" > .env

# Start everything
./scripts/start-all.sh

Service	URL
Dashboard	http://localhost:3000
API	http://localhost:8000
API Docs	http://localhost:8000/docs
Monitoring	http://localhost:8001

Logs are written to /tmp/meddocs/.

---

API

Patient & Tenant Management

Method	Endpoint	Description
POST	/api/tenants	Create tenant
POST	/api/tenants/signup	Create tenant + user
GET	/api/tenants/{id}/dashboard	Dashboard stats
POST	/api/tenants/{id}/patients	Create patient
GET	/api/tenants/{id}/patients	List patients

Document Processing

Method	Endpoint	Description
POST	/api/tenants/{id}/patients/{pid}/documents	Upload PDF/CSV
GET	/api/tenants/{id}/patients/{pid}/documents	List documents
GET	/api/tenants/{id}/documents/{did}	Document detail + extracted data
GET	/api/tenants/{id}/documents/{did}/file	View/download original file
DELETE	/api/tenants/{id}/documents/{did}	Delete document + data

Structured Data

Method	Endpoint	Description
GET	/api/tenants/{id}/patients/{pid}/lab-results	Lab results
GET	/api/tenants/{id}/patients/{pid}/prescriptions	Prescriptions
GET	/api/tenants/{id}/patients/{pid}/claims	Insurance claims
GET	/api/tenants/{id}/patients/{pid}/risk-flags	Risk flags
GET	/api/tenants/{id}/patients/{pid}/trends/{test}	Lab value trends

AI Features

Method	Endpoint	Description
POST	/api/tenants/{id}/patients/{pid}/query	Natural language search
GET	/api/tenants/{id}/patients/{pid}/summary	AI patient summary
POST	/api/tenants/{id}/patients/{pid}/reports/generate	Generate report
GET	/api/tenants/{id}/reports/{rid}/pdf	Download report PDF

Monitoring

Method	Endpoint	Description
GET	/monitor/tenants/{id}/active	Active pipeline runs
GET	/monitor/documents/{did}/status	Document processing status
GET	/monitor/documents/{did}/stream	SSE live events

---

Environment Variables

DATABASE_URL=postgresql+asyncpg://postgres:password@localhost:5433/meddocs
DATABASE_URL_SYNC=postgresql://postgres:password@localhost:5433/meddocs
REDIS_URL=redis://localhost:6379/0
GROQ_API_KEY=gsk_...
GROQ_REASONING_MODEL=llama-3.1-8b-instant
GROQ_CLASSIFICATION_MODEL=llama-3.1-8b-instant
EMBEDDING_MODEL=BAAI/bge-small-en-v1.5
EMBEDDING_DIM=384
STORAGE_PATH=./storage
MAX_UPLOAD_SIZE_MB=20

---

License

MIT