From Basic to Agentic RAG

This guide serves as the foundation for a workshop on Retrieval-Augmented Generation (RAG). You will learn how to build a RAG system using cycls, chromadb, and openai, starting with a basic implementation and evolving it into an “Agentic” RAG that handles file uploads and dynamic visualization tools. You will learn how to:

Build a Basic RAG system with static documents.
Create an Attachment Handler to process user uploads (PDFs, etc.).
Implement an Agentic RAG that indexes content dynamically.
Add Tooling capabilities to generate interactive charts.

Prerequisites

Python 3.8+
cycls package installed
OpenAI API Key
Docker installed (for local testing)

pip install cycls

Part 1: The Basic RAG

We start by creating a simple RAG system. This agent will have a pre-defined “knowledge base” hardcoded into it. It demonstrates the core loop of RAG: Embed -> Store -> Retrieve -> Generate. Create a file named basicrag.py.

Step 1.1: Setup and Dependencies

We initialize the agent with chromadb (vector database) and openai (embeddings and generation).

import cycls

# Initialize the agent with dependencies
agent = cycls.Agent(
    pip=["chromadb", "openai", "python-dotenv"],
    copy=[".env"]
)

Step 1.2: The Core Logic

The handler function performs the RAG operations. Note that in this basic version, documents are hardcoded.

@agent("chroma-agent", title="RAG Agent")
async def search_agent(context):
    import os
    from dotenv import load_dotenv
    import chromadb
    from chromadb.utils import embedding_functions

    load_dotenv()
    
    # 1. Setup OpenAI Embedding Function
    openai_ef = embedding_functions.OpenAIEmbeddingFunction(
        api_key=os.getenv("OPENAI_API_KEY"),
        model_name="text-embedding-3-small"
    )
    
    # 2. Initialize ChromaDB (Ephemeral in-memory for this demo)
    client = chromadb.Client()
    collection = client.get_or_create_collection(name="docs", embedding_function=openai_ef)

    # 3. Ingest Static Knowledge
    # In a real app, this would come from a database or file loader
    collection.add(
        documents=["I love cats", "I love dogs", "The weather is nice"],
        ids=["1", "2", "3"]
    )

    # 4. Retrieve Context
    query = context.messages[-1]["content"]
    results = collection.query(query_texts=[query], n_results=1)
    retrieved_doc = results['documents'][0][0]

    # 5. Generate Response (Simple yield for demo)
    yield f"Context: {retrieved_doc}"

# Run locally
agent.deploy(prod=False)

Why this is “Basic”:

The knowledge base is static and ephemeral (re-created on every run).
It cannot handle user files or new data.
It strictly retrieves text; it doesn’t do anything with it other than display it.

Part 2: Evolving to Agentic RAG

Now we move to agenticrag.py. An “Agentic” RAG doesn’t just look up info; it interacts with the environment. It will:

Read files you upload (PDFs, etc.).
Index them on the fly.
Decide whether to answer with text or generate a visualization (Tool Use).

Helper Module: `attach.py`

To keep our agent clean, we move complex file handling to attach.py. This module handles:

Downloading files from Cycls URLs.
Extracting text from PDFs (PyPDF2).
Formatting messages for the LLM.

Ensure attach.py is in the same directory.

Step 2.1: Agent Configuration

We need more dependencies now, including PyPDF2 for parsing and httpx for downloading.

agent = cycls.Agent(
    pip=["chromadb", "openai", "python-dotenv", "httpx", "PyPDF2"],
    copy=[".env", "attach.py"] # detailed copy instruction
)

Step 2.2: Dynamic Indexing

Instead of hardcoded strings, we process the incoming message to find file content.

# Inside agenticrag.py handler...

# Process Attachments using our helper
processed = await attach.process_messages_for_openai([context.messages[-1]])
last_msg_content = processed[0]['content']
full_text = last_msg_content if isinstance(last_msg_content, str) else last_msg_content[0]['text']

# Regex to find file content blocks formatted by attach.py
new_docs = re.findall(r'--- Content of .*? ---\n(.*?)\n--- End of file ---', full_text, re.DOTALL)

# Index new files immediately
if new_docs:
    try:
        collection.add(documents=new_docs, ids=[f"doc_{hash(d)}" for d in new_docs])
    except: pass

Step 2.3: Tool Use (Chart Generation)

This is what makes it “Agentic”. The model evaluates the user query. If the user asks for a “chart” or “plot”, it switches logic paths to generate HTML instead of just text.

# Define the tool
async def generate_chart(description: str, data: str) -> str:
    # ... (Prompt engineering to generate ApexCharts HTML) ...
    # See agenticrag.py for full implementation

# Decision Logic
if any(w in query.lower() for w in ["chart", "plot", "graph", "visualize"]):
    yield "Generating visualization..."
    chart_html = await generate_chart(query, retrieved_context)
    yield f'<iframe srcdoc="{chart_html}" ... ></iframe>'
else:
    # Standard RAG Text Response
    # ...

Key Differences Summary

Feature	Basic RAG (`basicrag.py`)	Agentic RAG (`agenticrag.py`)
Knowledge Base	Hardcoded, static strings.	Dynamic, built from user uploads.
Input Handling	Text only.	Text + Files (PDF, etc.) via `attach.py`.
Reasoning	Linear: Query -> Retrieve -> Answer.	Branching: Check intent -> (Visualize OR Answer).
Output	Plain Text.	Text or Interactive HTML Widgets.

Full Code Reference

1. Basic RAG (`basicrag.py`)

import cycls

# Initialize the agent with dependencies
agent = cycls.Agent(
    pip=["chromadb", "openai", "python-dotenv"],
    copy=[".env"]
)

@agent("chroma-agent", title="RAG Agent")
async def search_agent(context):
    import os
    from dotenv import load_dotenv
    import chromadb
    from chromadb.utils import embedding_functions

    load_dotenv()
    
    # Setup OpenAI Embedding Function
    openai_ef = embedding_functions.OpenAIEmbeddingFunction(
        api_key=os.getenv("OPENAI_API_KEY"),
        model_name="text-embedding-3-small"
    )
    
    # Initialize ChromaDB client
    client = chromadb.Client()
    
    # Create collection with the embedding function
    collection = client.get_or_create_collection(
        name="docs", 
        embedding_function=openai_ef
    )

    # Add documents to the collection
    collection.add(
        documents=["I love cats", "I love dogs", "The weather is nice"],
        ids=["1", "2", "3"]
    )

    # Query using the latest message
    query = context.messages[-1]["content"]
    results = collection.query(query_texts=[query], n_results=1)

    # Return retrieved context
    retrieved_doc = results['documents'][0][0]
    yield f"Context: {retrieved_doc}"

# Run locally
agent.deploy(prod=False)

2. Agentic RAG (`agenticrag.py`)

import cycls

# Initialize agent with dependencies for RAG + Attachments + Charts
agent = cycls.Agent(
    pip=["chromadb", "openai", "python-dotenv", "httpx", "PyPDF2"],
    copy=[".env", "attach.py"]
)

@agent("agentic-rag", title="Agentic RAG")
async def chat(context):
    import os
    import re
    from dotenv import load_dotenv
    import chromadb
    from chromadb.utils import embedding_functions
    from openai import OpenAI, AsyncOpenAI
    import attach 

    load_dotenv()
    
    # 1. Setup ChromaDB
    print("\n[DEBUG] 1. Setting up ChromaDB client and embedding function...")
    openai_ef = embedding_functions.OpenAIEmbeddingFunction(
        api_key=os.getenv("OPENAI_API_KEY"),
        model_name="text-embedding-3-small"
    )
    client = chromadb.Client()
    collection = client.get_or_create_collection(name="docs", embedding_function=openai_ef)
    print("[DEBUG] ChromaDB collection 'docs' ready.")

    # 2. Process Attachments
    print("[DEBUG] 2. Processing message for attachments...")
    processed = await attach.process_messages_for_openai([context.messages[-1]], debug=True)
    last_msg_content = processed[0]['content']
    full_text = last_msg_content if isinstance(last_msg_content, str) else last_msg_content[0]['text']
    
    # Index new files
    new_docs = re.findall(r'--- Content of .*? ---\n(.*?)\n--- End of file ---', full_text, re.DOTALL)
    if new_docs:
        print(f"[DEBUG] Found {len(new_docs)} new document sections to index.")
        try:
            collection.add(documents=new_docs, ids=[f"doc_{hash(d)}" for d in new_docs])
            print(f"[DEBUG] Successfully indexed {len(new_docs)} documents in ChromaDB.")
        except Exception as e:
            print(f"[DEBUG] Error indexing documents: {e}")
    else:
        print("[DEBUG] No new documents found to index.")

    # 3. Identify User Query
    original_content = context.messages[-1].get('content', '')
    query = ""
    if isinstance(original_content, str):
        query = original_content
    elif isinstance(original_content, list):
        query = " ".join([p['text'] for p in original_content if p.get('type') == 'text'])
    
    if not query.strip():
        query = "Summarize the uploaded document."
    
    print(f"[DEBUG] 3. User Query identified: '{query}'")

    # 4. RAG Retrieval
    print(f"[DEBUG] 4. Querying ChromaDB for top 1 result...")
    results = collection.query(query_texts=[query], n_results=1)
    retrieved_context = results['documents'][0][0] if results['documents'][0] else "No context found."
    print(f"[DEBUG] Retrieved Context length: {len(retrieved_context)} chars")
    print(f"[DEBUG] Context Preview: {retrieved_context[:100]}...")

    # --- Chart Generation Tool ---
    async def generate_chart(description: str, data: str) -> str:
        """Generates an HTML chart using ApexCharts via LLM."""
        vis_prompt = """
        You are a data visualization expert. Generate a single HTML file containing an interactive ApexCharts visualization.
        
        REQUIREMENTS:
        - Use the 'ApexCharts' library from CDN: <script src="https://cdn.jsdelivr.net/npm/apexcharts"></script>
        - The container div MUST have id="chart".
        - Initialize the chart in a <script> tag at the end of the body.
        - Use a modern, clean color palette (e.g., blues, teals, grays).
        - Ensure the chart is responsive (width: '100%').
        - Add tooltips and clear axis labels.
        - Title should be centered and styled appropriately.
        - Font family should be sans-serif (e.g., Inter, Roboto, Arial).
        
        Output ONLY the raw HTML code starting with <!DOCTYPE html>. Do not include markdown blocks (```html).
        Ensure the chart is beautiful, professional, and uses the provided data accurately.
        """
        llm_client = AsyncOpenAI(api_key=os.getenv("OPENAI_API_KEY")) # Re-using OpenAI for simplicity
        response = await llm_client.chat.completions.create(
            model="gpt-4o",
            messages=[
                {"role": "system", "content": vis_prompt},
                {"role": "user", "content": f"Description: {description}\nData: {data}"}
            ]
        )
        html = response.choices[0].message.content.strip()
        # Cleanup markdown if present
        if html.startswith("```"): html = html.split("\n", 1)[1].rsplit("\n", 1)[0]
        # Compress and escape for iframe
        compressed = " ".join(html.split()).replace('"', "&quot;")
        return compressed

    # 5. Decide: Chart or Text?
    if any(w in query.lower() for w in ["chart", "plot", "graph", "visualize"]):
        print("[DEBUG] 5. Decision: Generating visualization (Tool Use)")
        yield "Generating visualization..."
        chart_html = await generate_chart(query, retrieved_context)
        yield f'<iframe srcdoc="{chart_html}" width="100%" height="600" style="border:none; border-radius: 8px; box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.1); width: 100%; min-width: 600px;"></iframe>'
    else:
        print("[DEBUG] 5. Decision: Generating text response")
        # 6. Generate Text Answer
        llm = OpenAI()
        response = llm.chat.completions.create(
            model="gpt-4o",
            messages=[
                {"role": "system", "content": "You are a helpful RAG assistant. Answer using the provided context."},
                {"role": "user", "content": f"Context: {retrieved_context}\n\nQuestion: {query}"}
            ]
        )
        yield response.choices[0].message.content

# Run
agent.deploy(prod=False)

3. Attachment Handler (`attach.py`)

"""
Cycls Attachment Handler Module
================================
Handles attached files (PDF, Images, Text, CSV) and converts them to OpenAI format.

Usage:
---------
from attach import process_messages_for_openai, has_visual_content

async def llm(context):
    processed_messages = await process_messages_for_openai(context.messages)
    # ... use processed_messages with OpenAI
"""

import os
import re
import base64
import httpx


def extract_file_info(content_item):
    """Extract file URL and type from cycls file content item"""
    if content_item.get('type') != 'file':
        return None
    
    text = content_item.get('text', '')
    
    # Extract URL using regex
    url_match = re.search(r'File URL: (https?://[^\s\n]+)', text)
    type_match = re.search(r'File type: ([^\s\n]+)', text)
    name_match = re.search(r'\[File attached: ([^\]]+)\]', text)
    
    if url_match:
        return {
            'url': url_match.group(1),
            'type': type_match.group(1) if type_match else 'unknown',
            'name': name_match.group(1) if name_match else 'file'
        }
    return None


async def download_file(url, timeout=60.0):
    """Download file from URL and return bytes"""
    try:
        async with httpx.AsyncClient(timeout=timeout) as client:
            response = await client.get(url, follow_redirects=True)
            if response.status_code == 200:
                return response.content
            print(f"[attach] Download failed with status {response.status_code}")
    except Exception as e:
        print(f"[attach] Download error: {e}")
    return None


def extract_pdf_text(pdf_bytes):
    """Extract text from PDF bytes"""
    try:
        import PyPDF2  # type: ignore[import-untyped]
        import io
        
        pdf_file = io.BytesIO(pdf_bytes)
        pdf_reader = PyPDF2.PdfReader(pdf_file)
        
        text = ""
        for page in pdf_reader.pages:
            text += page.extract_text() + "\n"
        
        return text.strip()
    except Exception as e:
        print(f"[attach] PDF extraction error: {e}")
        return None


def get_mime_type(url_or_filename):
    """Detect mime type from URL or filename"""
    lower = url_or_filename.lower()
    if '.jpg' in lower or '.jpeg' in lower:
        return 'image/jpeg'
    elif '.gif' in lower:
        return 'image/gif'
    elif '.webp' in lower:
        return 'image/webp'
    elif '.png' in lower:
        return 'image/png'
    elif '.svg' in lower:
        return 'image/svg+xml'
    return 'image/png'  # default


async def process_messages_for_openai(messages, debug=False):
    """
    Process cycls messages and handle file attachments.
    
    Args:
        messages: List of message dicts from context.messages
        debug: If True, print debug information
    
    Returns:
        List of processed messages ready for OpenAI API
    """
    processed_messages = []
    
    for msg in messages:
        role = msg.get('role', 'user')
        content = msg.get('content', '')
        
        # If content is a string, keep it simple
        if isinstance(content, str):
            if content:  # Skip empty messages
                processed_messages.append({'role': role, 'content': content})
            continue
        
        # If content is a list, process each item
        if isinstance(content, list):
            new_content = []
            file_texts = []
            
            for item in content:
                item_type = item.get('type', '')
                
                if item_type == 'text':
                    text = item.get('text', '')
                    if text:
                        new_content.append({'type': 'text', 'text': text})
                
                elif item_type == 'image_url':
                    # Download image and convert to base64
                    image_url = item.get('image_url', {}).get('url', '')
                    
                    if image_url.startswith('data:'):
                        # Already base64, use as is
                        new_content.append(item)
                    elif image_url.startswith('http'):
                        if debug:
                            print(f"[attach] Downloading image...")
                        image_bytes = await download_file(image_url)
                        if image_bytes:
                            b64_data = base64.b64encode(image_bytes).decode('utf-8')
                            mime_type = get_mime_type(image_url)
                            new_content.append({
                                'type': 'image_url',
                                'image_url': {'url': f"data:{mime_type};base64,{b64_data}"}
                            })
                            if debug:
                                print(f"[attach] Image converted to base64 ({mime_type})")
                        else:
                            new_content.append({'type': 'text', 'text': '[Image could not be loaded]'})
                    else:
                        new_content.append(item)
                
                elif item_type == 'file':
                    # Extract file info and process
                    file_info = extract_file_info(item)
                    if file_info:
                        if debug:
                            print(f"[attach] Processing file: {file_info['name']} ({file_info['type']})")
                        
                        # Download the file
                        file_bytes = await download_file(file_info['url'])
                        
                        if file_bytes:
                            file_type = file_info['type'].lower()
                            
                            if 'pdf' in file_type:
                                # Extract text from PDF
                                pdf_text = extract_pdf_text(file_bytes)
                                if pdf_text:
                                    file_texts.append(f"\n\n--- Content of {file_info['name']} ---\n{pdf_text}\n--- End of file ---\n")
                                    if debug:
                                        print(f"[attach] Extracted {len(pdf_text)} chars from PDF")
                                else:
                                    file_texts.append(f"\n\n[Could not extract text from {file_info['name']}]\n")
                            
                            elif 'image' in file_type:
                                # Convert image to base64
                                b64_data = base64.b64encode(file_bytes).decode('utf-8')
                                mime_type = file_info['type']
                                new_content.append({
                                    'type': 'image_url',
                                    'image_url': {'url': f"data:{mime_type};base64,{b64_data}"}
                                })
                            
                            else:
                                # Try to read as text (txt, csv, json, etc.)
                                try:
                                    text_content = file_bytes.decode('utf-8')
                                    file_texts.append(f"\n\n--- Content of {file_info['name']} ---\n{text_content}\n--- End of file ---\n")
                                except:
                                    file_texts.append(f"\n\n[Could not read {file_info['name']}]\n")
            
            # Combine text content with file texts
            if file_texts:
                combined_text = ""
                for item in new_content:
                    if item.get('type') == 'text':
                        combined_text += item.get('text', '') + "\n"
                
                combined_text += "".join(file_texts)
                
                # If there are images, keep them and add combined text
                image_items = [item for item in new_content if item.get('type') == 'image_url']
                if image_items:
                    final_content = [{'type': 'text', 'text': combined_text.strip()}] + image_items
                    processed_messages.append({'role': role, 'content': final_content})
                else:
                    processed_messages.append({'role': role, 'content': combined_text.strip()})
            elif new_content:
                # No files, just regular content
                if len(new_content) == 1 and new_content[0].get('type') == 'text':
                    processed_messages.append({'role': role, 'content': new_content[0]['text']})
                else:
                    processed_messages.append({'role': role, 'content': new_content})
    
    return processed_messages


def has_visual_content(messages):
    """Check if processed messages contain images (need gpt-4o)"""
    return any(
        isinstance(msg.get('content'), list) and 
        any(item.get('type') == 'image_url' for item in msg.get('content', []))
        for msg in messages
    )

Get Started

Basics

Guide

Resources

LLMs

From Basic to Agentic RAG

Prerequisites

Part 1: The Basic RAG

Step 1.1: Setup and Dependencies

Step 1.2: The Core Logic

Part 2: Evolving to Agentic RAG

Helper Module: `attach.py`

Step 2.1: Agent Configuration

Step 2.2: Dynamic Indexing

Step 2.3: Tool Use (Chart Generation)

Key Differences Summary

Full Code Reference

1. Basic RAG (`basicrag.py`)

2. Agentic RAG (`agenticrag.py`)

3. Attachment Handler (`attach.py`)

Get Started

Basics

Guide

Resources

LLMs

​Prerequisites

​Part 1: The Basic RAG

​Step 1.1: Setup and Dependencies

​Step 1.2: The Core Logic

​Part 2: Evolving to Agentic RAG

​Helper Module: attach.py

​Step 2.1: Agent Configuration

​Step 2.2: Dynamic Indexing

​Step 2.3: Tool Use (Chart Generation)

​Key Differences Summary

​Full Code Reference

​1. Basic RAG (basicrag.py)

​2. Agentic RAG (agenticrag.py)

​3. Attachment Handler (attach.py)

Prerequisites

Part 1: The Basic RAG

Step 1.1: Setup and Dependencies

Step 1.2: The Core Logic

Part 2: Evolving to Agentic RAG

Helper Module: `attach.py`

Step 2.1: Agent Configuration

Step 2.2: Dynamic Indexing

Step 2.3: Tool Use (Chart Generation)

Key Differences Summary

Full Code Reference

1. Basic RAG (`basicrag.py`)

2. Agentic RAG (`agenticrag.py`)

3. Attachment Handler (`attach.py`)