Deploying AI Agents

This example demonstrates how to deploy an AI agent as a production HTTP service with an embedded web interface. The deployed agent analyzes documents and provides structured insights (summary, keywords, sentiment, readability) through both API and web UI.

Overview

The document analysis agent:

Ingests content from direct input, file uploads, or URLs
Extracts structured insights: summary, keywords, sentiment, readability
Runs online or offline: Uses OpenAI if API key is set, otherwise fallback
Provides HTTP API: RESTful endpoint for programmatic access
Includes web UI: Modern SPA interface embedded in the deployment
Returns HTML reports: Visualization for the ZenML dashboard

Source code

The complete example is available at:

https://github.com/zenml-io/zenml/tree/main/examples/deploying_agent

Quick start

Installation

git clone https://github.com/zenml-io/zenml.git
cd zenml/examples/deploying_agent

# Install dependencies
pip install -r requirements.txt

# Optional: Set OpenAI API key for LLM analysis
export OPENAI_API_KEY=sk-xxx

# Initialize ZenML
zenml init
zenml login

Deploy the pipeline

# Deploy as HTTP service
zenml pipeline deploy pipelines.doc_analyzer.doc_analyzer

# Get endpoint URL
zenml deployment describe doc_analyzer

Invoke the agent

Via CLI

zenml deployment invoke doc_analyzer \
  --content="Artificial Intelligence is transforming how we work..." \
  --filename="ai-overview.txt" \
  --document_type="text"

Via HTTP API

curl -X POST http://localhost:8000/invoke \
  -H "Content-Type: application/json" \
  -d '{
    "parameters": {
      "content": "Your document content here...",
      "filename": "document.txt",
      "document_type": "text"
    }
  }'

Via web interface

Visit the deployment URL in your browser:

http://localhost:8000

The web UI provides three input methods:

Direct Content: Paste or type content directly
Upload File: Upload text files, markdown, or HTML
URL: Analyze content from a URL

Pipeline structure

from zenml import pipeline, ArtifactConfig
from zenml.config import DockerSettings, DeploymentSettings, CORSConfig
from typing import Annotated, Optional
from models import DocumentAnalysis

docker_settings = DockerSettings(
    requirements="requirements.txt",
    environment={
        "OPENAI_API_KEY": "${OPENAI_API_KEY}",
    },
)

deployment_settings = DeploymentSettings(
    app_title="Document Analysis Pipeline",
    dashboard_files_path="ui",  # Serve web UI from ui/ directory
    cors=CORSConfig(allow_origins=["*"]),  # Enable CORS for web access
)

@pipeline(
    settings={
        "docker": docker_settings,
        "deployment": deployment_settings,
    },
    enable_cache=False,  # Disable caching for real-time serving
)
def doc_analyzer(
    content: Optional[str] = None,
    url: Optional[str] = None,
    path: Optional[str] = None,
    filename: Optional[str] = None,
    document_type: str = "text",
) -> Annotated[
    DocumentAnalysis,
    ArtifactConfig(name="document_analysis", tags=["analysis", "serving"]),
]:
    """Document analysis pipeline deployed as HTTP service.
    
    Args:
        content: Direct text content (optional)
        url: URL to download content from (optional)
        path: Path to file (optional)
        filename: Document name (auto-generated if not provided)
        document_type: Type of document (text, markdown, report, article)
    
    Returns:
        DocumentAnalysis: Complete analysis results
    """
    # Ingest document from various sources
    document = ingest_document_step(
        content=content,
        url=url,
        path=path,
        filename=filename,
        document_type=document_type,
    )
    
    # Analyze document (LLM or deterministic fallback)
    analysis = analyze_document_step(document)
    
    # Generate HTML report for dashboard
    render_analysis_report_step(analysis)
    
    return analysis

Pipeline steps

1. Document ingestion

@step
def ingest_document_step(
    content: Optional[str] = None,
    url: Optional[str] = None,
    path: Optional[str] = None,
    filename: Optional[str] = None,
    document_type: str = "text",
) -> Annotated[DocumentRequest, "document"]:
    """Ingest document from various sources.
    
    Supports three ingestion modes:
    1. Direct content: Pass text directly
    2. URL: Download from web (with HTML cleaning)
    3. Path: Load from file system or artifact store
    """
    if content:
        # Direct content ingestion
        doc_content = content
        doc_filename = filename or f"document_{int(time.time())}.txt"
    
    elif url:
        # Download from URL
        import requests
        from bs4 import BeautifulSoup
        
        response = requests.get(url, timeout=30)
        response.raise_for_status()
        
        # Clean HTML if needed
        if "text/html" in response.headers.get("content-type", ""):
            soup = BeautifulSoup(response.content, "html.parser")
            doc_content = soup.get_text(separator="\n", strip=True)
        else:
            doc_content = response.text
        
        doc_filename = filename or url.split("/")[-1] or "document.txt"
    
    elif path:
        # Load from file
        with open(path, "r") as f:
            doc_content = f.read()
        doc_filename = filename or os.path.basename(path)
    
    else:
        raise ValueError("Must provide content, url, or path")
    
    # Validate content
    if not doc_content or not doc_content.strip():
        raise ValueError("Document content is empty")
    
    return DocumentRequest(
        filename=doc_filename,
        content=doc_content,
        document_type=document_type,
        word_count=len(doc_content.split()),
    )

2. Document analysis

@step
def analyze_document_step(
    document: DocumentRequest,
) -> Annotated[DocumentAnalysis, "document_analysis"]:
    """Analyze document using LLM or deterministic fallback.
    
    Attempts LLM analysis first (OpenAI), falls back to rule-based
    analysis if LLM is unavailable.
    """
    # Validate input
    if not document.content or not document.content.strip():
        raise ValueError(f"Empty document: {document.filename}")
    
    # Try LLM analysis
    try:
        analysis_result = perform_llm_analysis(
            content=document.content,
            filename=document.filename,
        )
        analysis_method = "llm"
        model_label = f"AI ({analysis_result['used_model']})"
    
    except Exception:
        # Fallback to deterministic analysis
        analysis_result = perform_deterministic_analysis(
            content=document.content,
            filename=document.filename,
        )
        analysis_method = "deterministic_fallback"
        model_label = "rule-based (deterministic)"
    
    # Create analysis object
    analysis = DocumentAnalysis(
        document=document,
        summary=analysis_result["summary"],
        keywords=analysis_result["keywords"],
        sentiment=analysis_result["sentiment"],
        word_count=len(document.content.split()),
        readability_score=analysis_result["readability_score"],
        model=model_label,
        latency_ms=analysis_result["latency_ms"],
        tokens_prompt=analysis_result["tokens_prompt"],
        tokens_completion=analysis_result["tokens_completion"],
        metadata={
            "source": "document_analysis_pipeline",
            "analysis_method": analysis_method,
            "document_type": document.document_type,
        },
    )
    
    return analysis

3. Report generation

@step
def render_analysis_report_step(
    analysis: DocumentAnalysis,
) -> Annotated[str, "analysis_report"]:
    """Generate HTML report for dashboard visualization."""
    html = f"""<!DOCTYPE html>
<html>
<head>
    <title>Document Analysis: {analysis.document.filename}</title>
    <link rel="stylesheet" href="report.css">
</head>
<body>
    <div class="container">
        <h1>Document Analysis Report</h1>
        
        <div class="section">
            <h2>Document Information</h2>
            <p><strong>Filename:</strong> {analysis.document.filename}</p>
            <p><strong>Type:</strong> {analysis.document.document_type}</p>
            <p><strong>Word Count:</strong> {analysis.word_count:,}</p>
        </div>
        
        <div class="section">
            <h2>Summary</h2>
            <p>{analysis.summary}</p>
        </div>
        
        <div class="section">
            <h2>Key Metrics</h2>
            <div class="metrics">
                <div class="metric">
                    <h3>Sentiment</h3>
                    <p class="sentiment-{analysis.sentiment}">{analysis.sentiment.title()}</p>
                </div>
                <div class="metric">
                    <h3>Readability</h3>
                    <p>{analysis.readability_score:.2f}</p>
                </div>
                <div class="metric">
                    <h3>Processing Time</h3>
                    <p>{analysis.latency_ms}ms</p>
                </div>
            </div>
        </div>
        
        <div class="section">
            <h2>Keywords</h2>
            <div class="keywords">
                {''.join(f'<span class="keyword">{kw}</span>' for kw in analysis.keywords)}
            </div>
        </div>
        
        <div class="section">
            <h2>Analysis Details</h2>
            <p><strong>Model:</strong> {analysis.model}</p>
            <p><strong>Tokens (prompt):</strong> {analysis.tokens_prompt:,}</p>
            <p><strong>Tokens (completion):</strong> {analysis.tokens_completion:,}</p>
        </div>
    </div>
</body>
</html>"""
    
    return html

LLM analysis implementation

def perform_llm_analysis(
    content: str,
    filename: str,
    model: str = "gpt-4o-mini",
) -> Dict[str, Any]:
    """Perform document analysis using OpenAI."""
    from openai import OpenAI
    import json
    
    # Clean and truncate content
    cleaned_content = clean_text_content(content)
    content_preview = cleaned_content[:4000]  # First 4000 chars
    
    # Build analysis prompt
    prompt = f"""Analyze the following document and provide:
1. A concise summary (2-3 sentences)
2. Top 5 keywords
3. Sentiment (positive/negative/neutral)
4. Readability level (easy/medium/hard)

Document: {filename}

Content:
{content_preview}

Respond with JSON:
{{
  "summary": "...",
  "keywords": ["...", "...", "...", "...", "..."],
  "sentiment": "...",
  "readability": "..."
}}"""
    
    start_time = time.time()
    
    # Call OpenAI API
    client = OpenAI()
    response = client.chat.completions.create(
        model=model,
        messages=[
            {"role": "system", "content": "You are a document analysis expert."},
            {"role": "user", "content": prompt},
        ],
        max_tokens=500,
        temperature=0.3,  # Low temperature for consistency
    )
    
    latency_ms = int((time.time() - start_time) * 1000)
    
    # Parse JSON response
    response_text = response.choices[0].message.content
    analysis_response = json.loads(response_text)
    
    # Map readability to score
    readability_map = {"easy": 0.8, "medium": 0.5, "hard": 0.3}
    readability_score = readability_map.get(
        analysis_response.get("readability", "medium").lower(),
        0.5,
    )
    
    return {
        "summary": analysis_response["summary"],
        "keywords": analysis_response["keywords"][:5],
        "sentiment": analysis_response["sentiment"],
        "readability_score": readability_score,
        "tokens_prompt": response.usage.prompt_tokens,
        "tokens_completion": response.usage.completion_tokens,
        "latency_ms": latency_ms,
        "used_model": model,
    }

Deterministic fallback

def perform_deterministic_analysis(
    content: str,
    filename: str,
) -> Dict[str, Any]:
    """Rule-based analysis when LLM is unavailable."""
    from collections import Counter
    
    start_time = time.time()
    
    # Extract summary from first paragraph
    paragraphs = content.split("\n\n")
    summary = paragraphs[0][:200] + "..." if len(paragraphs[0]) > 200 else paragraphs[0]
    
    # Simple keyword extraction
    words = clean_text_content(content).lower().split()
    stop_words = {"the", "a", "an", "and", "or", "but", "in", "on", "at", "to", "for"}
    
    filtered_words = [
        w for w in words
        if len(w) > 3 and w not in stop_words and w.isalpha()
    ]
    
    word_freq = Counter(filtered_words)
    keywords = [word for word, _ in word_freq.most_common(5)]
    
    # Ensure 5 keywords
    while len(keywords) < 5:
        keywords.append(f"keyword{len(keywords) + 1}")
    
    # Default sentiment
    sentiment = "neutral"
    
    # Readability based on average word length
    avg_word_len = sum(len(w) for w in words) / len(words) if words else 5
    readability_score = max(0.1, 1.0 - (avg_word_len - 4) / 10)
    
    latency_ms = int((time.time() - start_time) * 1000)
    
    return {
        "summary": summary,
        "keywords": keywords,
        "sentiment": sentiment,
        "readability_score": readability_score,
        "tokens_prompt": len(content.split()),
        "tokens_completion": len(summary.split()),
        "latency_ms": latency_ms,
    }

Web UI

The embedded web interface (ui/index.html) provides:

Multi-tab interface: Direct content, file upload, or URL analysis
Real-time feedback: Loading states and error messages
Results display: Summary, sentiment, keywords, metrics
Responsive design: Works on desktop and mobile
Zero configuration: Automatically served at deployment URL

The UI is configured via DeploymentSettings:

deployment_settings = DeploymentSettings(
    app_title="Document Analysis Pipeline",
    dashboard_files_path="ui",  # Serve files from ui/ directory
    cors=CORSConfig(allow_origins=["*"]),  # Enable CORS
)

Deployment configuration

Docker settings

docker_settings = DockerSettings(
    requirements="requirements.txt",
    python_package_installer="uv",  # Fast installs
    environment={
        "OPENAI_API_KEY": "${OPENAI_API_KEY}",  # Pass from host
    },
)

Custom configuration

Create a YAML config for advanced settings:

# deployment_config.yaml
settings:
  deployer:
    generate_auth_key: true  # Enable authentication
    resources:
      cpu: "2"
      memory: "4Gi"

Deploy with config:

zenml pipeline deploy pipelines.doc_analyzer.doc_analyzer \
  --config deployment_config.yaml

Production considerations

Authentication: Enable generate_auth_key for production
Rate limiting: Implement request throttling
Monitoring: Track latency, errors, and token usage
Scaling: Configure replica count for high traffic
Costs: Monitor OpenAI API usage and costs
Fallback: Ensure deterministic analysis works without API key
Error handling: Return user-friendly error messages
Validation: Sanitize and validate all inputs

Testing the deployment

Health check

curl http://localhost:8000/health

Test analysis

curl -X POST http://localhost:8000/invoke \
  -H "Content-Type: application/json" \
  -d '{
    "parameters": {
      "content": "Machine learning is revolutionizing software development. AI models can now understand context, generate code, and assist developers in ways that were impossible just a few years ago. This technology is making software development more accessible and efficient.",
      "filename": "ai-ml-overview.txt",
      "document_type": "article"
    }
  }'

Expected response

{
  "document": {
    "filename": "ai-ml-overview.txt",
    "content": "Machine learning is revolutionizing...",
    "document_type": "article",
    "word_count": 42
  },
  "summary": "Machine learning and AI are transforming software development by enabling new capabilities in code generation and developer assistance.",
  "keywords": ["machine", "learning", "software", "development", "models"],
  "sentiment": "positive",
  "readability_score": 0.65,
  "model": "AI (gpt-4o-mini)",
  "latency_ms": 847,
  "tokens_prompt": 156,
  "tokens_completion": 32
}

Next steps

Agent comparison

Compare multiple agent architectures systematically

Framework integrations

Examples for 12+ agent frameworks

Orchestrating agents

Production orchestration patterns

Agent evaluation

Build systematic evaluation pipelines

Agent Workflows

Agent Examples

Deploying AI Agents

Overview

Source code

Quick start

Installation

Deploy the pipeline

Invoke the agent

Via CLI

Via HTTP API

Via web interface

Pipeline structure

Pipeline steps

1. Document ingestion

2. Document analysis

3. Report generation

LLM analysis implementation

Deterministic fallback

Web UI

Deployment configuration

Docker settings

Custom configuration

Production considerations

Testing the deployment

Health check

Test analysis

Expected response

Next steps

Agent comparison

Framework integrations

Orchestrating agents

Agent evaluation

Build docs developers (and LLMs) love

Agent Workflows

Agent Examples

​Overview

​Source code

​Quick start

​Installation

​Deploy the pipeline

​Invoke the agent

​Via CLI

​Via HTTP API

​Via web interface

​Pipeline structure

​Pipeline steps

​1. Document ingestion

​2. Document analysis

​3. Report generation

​LLM analysis implementation

​Deterministic fallback

​Web UI

​Deployment configuration

​Docker settings

​Custom configuration

​Production considerations

​Testing the deployment

​Health check

​Test analysis

​Expected response

​Next steps

Agent comparison

Framework integrations

Orchestrating agents

Agent evaluation

Build docs developers (and LLMs) love

Overview

Source code

Quick start

Installation

Deploy the pipeline

Invoke the agent

Via CLI

Via HTTP API

Via web interface

Pipeline structure

Pipeline steps

1. Document ingestion

2. Document analysis

3. Report generation

LLM analysis implementation

Deterministic fallback

Web UI

Deployment configuration

Docker settings

Custom configuration

Production considerations

Testing the deployment

Health check

Test analysis

Expected response

Next steps