Webinar: Building an AI-Ready Data Backbone with Aiven & Google Cloud

“LLM became an operating system: it does not do the work by itself, but it can reason about what it sees.”

This past Thursday, 26 March, I had a pleasure to deliver a webinar by invitation of the AI Camp. The title of the webinar was “Building an AI-ready Data Backbone”, with the focus on how using OpenSearch you can implement a more intelligent search experience, than pure keyword retrieval. The demo I showed (and have it open sourced) implemented an agentic search over the episodes of the Vector Podcast.

Top slide for my deck

The demo allows to handle intricate questions like: “I’m new to Vector Podcast and vector search. Build a beginner-to-advanced learning path from episodes.” And the system searches episodes, summarizes and analyzes them and then presents a learning path for this request:

System response for a question classified as a “learning path” intent

No matter what agentic workflow you have implemented, the key bit is to establish a feedback loop where you can start with analyzing the system output for quality and recording this per intent.

Quality dashboard with user feedback, grouped by intent

I’m currently focusing on the managed OpenSearch service at Aiven, which is praised by clients as “as easy as a button click”.

You might be wondering why OpenSearch vs other engines on the market:

• OpenSearch is a truly open source governed search engine, being part of the Linux Foundation
• There are a number of industry players that shape its roadmap: Amazon, Uber, Apple, OpenSource Connections, Aiven
• It stands on battle tested Apache Lucene — search library with a vast usage across the industry and research, with over 25 years in existence!
• The feature set you care for when building modern search products, like multimodal vector / similarity search, aggregations (or facets — that allows you to display counts next to brands, categories, sizes, colours etc), geospatial search support and of course full-text search are all included.
• Add to these the powerful OpenSearch Dashboards, for querying and visualization of your data, running query labelling, and eye-balling your query performance:

Single query comparison view in Search Relevance Workbench section of OpenSearch Dashboards

Here is the recording:

[embed]

We have covered a few topics:

🔹 OpenSearch: The Foundation for AI

OpenSearch has evolved far beyond a simple fork of Elasticsearch. It is now a critical engine for:

• Vector & Multimodal Search: Essential for building RAG (Retrieval-Augmented Generation) systems.
• Hybrid Retrieval: Combining traditional keyword search with vector-based neural search for precision.
• Scalability: Aiven’s new Dedicated Node Roles (Cluster Manager vs. Data Nodes) allow for high-performance scaling at a lower cost.

🔹 From Vanilla RAG to Agentic RAG

In the webinar I discussed a major shift from “static” RAG pipelines to Agentic RAG. Unlike standard systems, AI Agents can:

• Self-Critique: Evaluate retrieved context and decide if it needs to iterate.
• Use Tools: Dynamically choose between vector search, web search, or APIs.
• Maintain Memory: Track both short-term conversation state and long-term user preferences.

AI Agent components

One “trick” I did to melt the ice with the audience was to invite everyone to interrupt me at any point and ask their questions. Here is a sample of questions (you will find them all in the recording below):

1. How does the OpenSearch RAG handle conversation history to maintain context across multiple follow-up questions?
2. How does you agentic memory implementation handles state consistency without creating a massive context window? Are you using a shared Kafka topic for transient state or a unified Postgres SQL check-pointer for procedural memory?
3. Can I use Kafka to audit agentic conversation history?
4. (In RAG or agentic search application) How do you determine that you did not get a good enough answer?
5. Shouldn’t you be building evals at every step of your application instead of just doing the final evaluation of the final outcome?
6. On smaller scale human-in-the-loop (HILT — abbreviation proposed by an audience member!) kinda makes sense for assessing trust towards the quality, but how does this work at scale?
7. To support autonomous decision-making of AI agents, does Aiven now support MCP?

Our discussion with the audience also touched on graph methods for RAG, as opposed to more linear flows of searching for answers. Think of the query “List all companies acquired by IBM between 2006 and 2026”. All exciting topics, and it feels that they can only resurface in a live conversation like this!

🎁 Big News: Free Tier OpenSearch on Aiven!

Aiven has launched a completely free tier for OpenSearch on their platform. It’s a dedicated service with 4GB RAM and 20GB storage — perfect for builders, learners, and prototypers to start indexing in seconds with no credit card required. Access it here: https://console.aiven.io/signup?campaign=opensearch_free_tier

Want to dive into the code? Check out the workshop repo for building an Agentic Search UI using Streamlit and Aiven OpenSearch: 🔗

Application: Agentic Search

Data: Vector Podcast episodes

UI: Streamlit

Search backend: Aiven OpenSearch v3.3

1. Episode download / transcribe / index: https://github.com/dimakan-dev/conduit-transcripts/blob/main/DATA_PROCESSING_GUIDE.md
2. Main UI and quality dashboards: https://github.com/dimakan-dev/preparing-data-for-opensearch-and-rag/blob/main/workshop/STREAMLIT_README.md