Pondhouse Data AI - Tips & Tutorials for Data & AI 19

Subscribe, if this was forwarded to you

Hey there,

This week we're focusing on practical AI developments that matter for developers and businesses: The EU AI Act's first phase is now active, with clear guidelines on prohibited applications. We've got a detailed tutorial on building RAG systems with Azure AI Search, and we're exploring n8n, an open-source platform that makes AI agent development surprisingly accessible.

In our testing section, we've been impressed by Google's Gemini 2.0, particularly its document processing capabilities, and we're sharing a great find for local text-to-speech: Kokoro, a lightweight but powerful open-source model that's perfect for production environments.

Enjoy the read!

Cheers, Andreas & Sascha

In todays edition:

📚 Tutorial of the Week: Building Production-Ready RAG Systems with Azure AI Search

🛠️ Tool Spotlight: n8n - The Open-Source Platform for Building AI Agents

📰 Top News:

• EU AI Act Takes Effect: What Companies Need to Know

• OpenAI Releases Deep Research for Autonomous Web Research

• Google's Gemini 2.0 Launches with Three Variants

• Global AI Summit Shows Shifting Industry Dynamics

💡 Tip of the Week: Local Text-to-Speech with Kokoro - High Quality, Low Resources

Let's get started!

Find this Newsletter helpful?
Please forward it to your colleagues and friends - it helps us tremendously.

Tutorial of the week

Building a RAG System with Azure AI Search

While Large Language Models are powerful, they need accurate, up-to-date information to provide reliable responses. Retrieval-Augmented Generation (RAG) solves this by feeding relevant context to LLMs before they generate answers. Azure AI Search provides an enterprise-ready solution for the crucial retrieval part of RAG, making it a great choice for teams looking to build production-ready RAG systems.

Key Features of Azure AI Search for RAG:

• Vector search for semantic understanding (finds relevant content even when exact keywords don't match)

• Hybrid search combining traditional keywords and vectors for better accuracy

• Automatic data ingestion from various sources (PDFs, Word docs, databases)

• Built-in AI enrichment capabilities for data processing

• Enterprise-grade security and scalability

• Managed service requiring minimal maintenance

Quick-Start Guide:

1. Set Up Azure AI Search:

   - Go to Azure Portal - Create new Azure AI Search service - Select "Free" tier for testing - This gives you 50MB storage and basic features to experiment

2. Prepare your data
   - Upload documents to Azure Blob Storage - Create a data source in Azure AI Search - Use the import wizard for initial setup - Azure will automatically extract text and metadata

3. Configure your index
   - Select embedding model (e.g., OpenAI's ada-002) - Define searchable fields (content, metadata, etc.) - Configure vector search settings - Set up indexing schedule for automatic updates

4. Integration options
   A. Azure AI Studio (Easiest Start): - Create new project - Connect Azure AI Search - Use chat playground for immediate testing - Perfect for prototyping and quick demos B. REST API (Production Use): - Get endpoint and API keys - Use provided curl command template - Integrate into your application - Offers more flexibility and control

Tips:

• Start with small document sets to understand the system

• Use hybrid search to get the best of both vector and keyword matching

• Monitor your index size and adjust chunking strategy accordingly

• Consider semantic ranking for better result ordering (available in paid tiers)

Pitfalls to avoid:

• Don't skip proper document chunking - it's crucial for good results

• Remember to handle document updates and deletions

• Test with various query types to ensure robust performance

• Monitor API usage to stay within limits

For a full tutorial, read more in our full-length article here.

Tool of the week

n8n - The best Open-Source, Low-Code AI Agent Builder

n8n is an open-source workflow automation platform that has evolved into a powerful tool for building AI agents. Think of it as a visual programming environment where you can create complex AI workflows without writing code - though you can add code when needed (and trust us, each and every medium-sized low-code project will eventually need some coding extension!).

Why It Matters Now

With the explosion of AI services and the growing need for custom AI agents, n8n hits a sweet spot between accessibility and usefuleness. It allows you to visually construct AI workflows connecting multiple services like OpenAI, Anthropic, and custom APIs, while adding components like memory management and reasoning frameworks. Being open-source and self-hostable makes it also very attractive for teams working with sensitive data or requiring full control over their infrastructure.

Key features

• Visual workflow builder

• JavaScript code nodes for custom logic, including npm support

• Error handling and retries built in

• Version control integration

• More than 500 APIs supported

• Native LLM integrations for many providers

• Memory management nodes for agents

• Agent reasoning frameworks

• Webhook support for realtime responses

Why you’ll love n8n for AI agent creation

What sets n8n apart is its thoughtful approach to AI agent development. You start with an intuitive visual interface where you can drag and drop components to create your agent's logic. Why are we - as deep-tech-enthusiasts - fancying ourselves with drag and drop logic? Because for AI agents you desperately and always need to sketch your architecture and dataflows before implementing them. Using n8n basically allows you to sketch and implement at the same time!
Need to add memory to your chatbot? There's a node for that. Want to implement chain-of-thought reasoning? You can visual craft the logic flow. When you inevitably need more flexibility, the JavaScript code nodes let you add custom logic without leaving the platform - even allowing to pull npm packages.

Community and Resources

One of n8n's other strong assets is its vibrant community. With over 1,200 pre-made workflow templates available for free, you're never starting from scratch. The community regularly shares new AI patterns, solutions, and creative implementations. The official documentation is comprehensive, and the team or community regularly host webinars and tutorials focused on AI implementation.

Getting started

Starting with n8n is straightforward: you can either use their cloud offering or self-host the platform. Begin by exploring some AI agent templates to understand the possibilities, then start building your own workflows. Connect your preferred AI services, add logic nodes, and you're on your way to creating your first agent.

Pro Tips:

• Start simple but plan for complexity - your agents will grow more sophisticated

• Don't shy away from the code nodes; they're your friends when the visual interface hits its limits

• Use the community templates to learn best practices

• Consider self-hosting if data privacy is a concern

The learning curve is gentle but deep: while you can create basic workflows within minutes, mastering AI agent development takes time and experimentation - while the superb user interface makes experimentation a joy.

Top News of the week

EU AI Act Takes Effect: What companies need to know

The EU made history last year by introducing the first AI regulation framework for a global jurisdiction.

The first phase of the EU AI Act is now active, focusing on prohibiting specific AI applications deemed too risky for society. While full compliance isn't required until mid-2025, companies must already avoid certain practices or face severe penalties.

Key Prohibited Applications:

• Social scoring systems

• Emotion recognition in workplaces and schools

• Real-time biometric identification in public spaces

• Untargeted facial recognition database creation

• Manipulative or exploitative AI systems

Companies violating these regulations face penalties of up to 7% of their global annual turnover - a figure that surpasses even GDPR penalties. This sends a clear message about the EU's ideas for responsible AI development.

The regulations affect any company serving EU customers or using AI systems that impact EU citizens, regardless of where the company is based.

What you need to do:

First, are you implementing a social scoring system, emotion recognition, biometric identification, facial recognition database or plan to use AI for manipulating clients? If not, there is currently nothing to worry about. For most parts of the economy, this regulation does not limit development and use of AI - this first phase is clearly targeted to consumer protection.

However, while the initial prohibitions are now in effect, the more comprehensive requirements are coming in waves. Here's how organizations should prepare for the next phases:

Timeline Overview:

• February 2025: Initial prohibitions and AI literacy requirements

• August 2025: GPAI model rules and governance structures

• August 2026: Full application of the Act

• August 2027: Extended deadline for high-risk AI systems in regulated products

Recommended Immediate Actions (Next 6 Months):

1. AI Inventory
   Create a comprehensive catalog of all AI systems in use, categorizing them by risk level according to the Act's framework. This is your foundation for all future compliance work (and might also help to get your AI projects in order)

2. Focus on AI Literacy
   The Act specifically requires organizations to ensure staff have "sufficient AI literacy." Start developing:

   • Training programs for different roles

   • Documentation of AI systems

   • Clear guidelines for AI use

   • Regular assessment of AI competency

3. Implement basic AI governance

   1. Dedicated AI oversight role

   2. Create internal guidelines for AI use

IMPORTANT: Transparency risks

While not in effect yet, this part of the act will roll out next: AI Transparency.

The AI Act introduces specific disclosure obligations to ensure that humans are informed when necessary to preserve trust. For instance, when using AI systems such as chatbots, humans should be made aware that they are interacting with a machine so they can take an informed decision.

Moreover, providers of generative AI have to ensure that AI-generated content is identifiable. On top of that, certain AI-generated content should be clearly and visibly labelled, namely deep fakes and text published with the purpose to inform the public on matters of public interest.

To get the official AI Act summary from the EU, click here.

OpenAI releases Deep Research - a way to automate web research

OpenAI has released "deep research," a significant upgrade to ChatGPT that enables autonomous, multi-step research by intelligently browsing and analyzing online sources. The system achieved 26.6% accuracy on "Humanity's Last Exam" benchmark (previous best was 13%) and can save hours on complex research tasks by intelligently searching, verifying facts, and citing sources.

The company plans combining deep research with their "Operator" feature for complete task execution.

Read more here.

Google Releases Gemini 2.0

Google has made its highly anticipated Gemini 2.0 generally available, introducing three variants: Flash (the workhorse model with 1M token context), Pro (their most advanced model with 2M token context), and Flash-Lite (their most cost-efficient option). The standout feature is Flash's ability to handle high-volume, high-frequency tasks at scale with impressive multimodal reasoning capabilities. Pro version excels at coding and complex prompts, while Flash-Lite offers better quality than its predecessor at the same speed and cost.

OUR TESTING: We've tested Gemini 2.0 for document intelligence tasks, and the results are remarkable. The system demonstrates exceptional accuracy in processing and analyzing documents, delivering results significantly faster than competitors. Most impressively, it maintains a highly competitive price point, being up to 100x cheaper than OpenAIs offerings, making it an excellent choice for businesses requiring large-scale document processing capabilities.

Read more about the model here.

Global AI Summit Shows Shifting Industry Dynamics

Representatives from 80 countries are gathering in Paris for a two-day AI summit, with discussions focused on safety standards, international collaboration, and industry regulation. The meeting comes as China's DeepSeek has demonstrated significant technical achievements, challenging assumptions about U.S. companies' technical lead in AI development. Key attendees include U.S. Vice President JD Vance, OpenAI's Sam Altman, Google's Sundar Pichai, and China's senior leader Ding Xuexiang.

The summit aims to address practical challenges in AI development and deployment, including safety protocols, responsible innovation, and international standards. While previous summits focused heavily on existential risks, this meeting is expected to tackle immediate concerns such as misinformation, bias in AI systems, and the need for transparent development practices. European leaders, including French President Macron, are positioning the EU to play a larger role in shaping global AI policy, emphasizing the need for balanced international cooperation in AI advancement.

WHY IT MATTERS: As AI technology becomes increasingly central to global economic and technological development, international cooperation and standard-setting become essential for ensuring responsible innovation and fair competition.

Tip of the week

Local Text-To-Speech with Kokoro

Kokoro is an impressive new open-source text-to-speech model that's making waves in the AI community. With just 82 million parameters (tiny compared to most modern AI models), it delivers remarkably high-quality voice synthesis across 8 languages and 54 voices. What makes it special is its combination of speed, quality, and ease of use - all while being completely free and Apache-licensed.

Using it could not be easier:

# 1. Install Kokoro
!pip install kokoro soundfile

# 2. Install espeak for language support
!apt-get install espeak-ng  # Linux/WSL
# For Windows, download from: https://github.com/espeak-ng/espeak-ng

# 3. Basic usage
from kokoro import KPipeline
import soundfile
pipeline = KPipeline(lang_code='a')  # 'a' for American English

# 4. Generate speech
text = "Hello, this is a test of Kokoro TTS!"
generator = pipeline(text, voice='af_heart', speed=1)
for _, _, audio in generator:
    # Save to WAV file
    soundfile.write('output.wav', audio, 24000)

TIPS:

• Use lang_code to switch languages ('a' for US English, 'b' for UK English, 'j' for Japanese, etc.)

• Adjust speed parameter between 0.5-2.0 to control speech rate

• Check VOICES.md in the GitHub repo for all available voices

• For even better performance, run on GPU if available

Check out some samplese here.

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help