A very common question I get when presenting and talking about advanced RAG (Retrieval Augmented Generation) techniques, is how to best index and search rich documents like PDF (or web pages), that contain both text and rich elements, like pictures or diagrams.

Another very frequent question that people ask me is about RAG versus long context windows. Indeed, models with long context windows usually have a more global understanding of a document, and each excerpt in its overall context. But of course, you can’t feed all the documents of your users or customers in one single augmented prompt. Also, RAG has other advantages like offering a much lower latency, and is generally cheaper.

In the first article of this Advanced RAG series, I talked about an approach I called sentence window retrieval, where we calculate vector embeddings per sentence, but the chunk of text returned (and added in the context of the LLM) actually contains also surrounding sentences to add more context to that embedded sentence. This tends to give a better vector similarity than the whole surrounding context. It is one of the techniques I’m covering in my talk on advanced RAG techniques.

When I talk about Generative AI, whether it’s with developers at conferences or with customers, I often find myself saying the same thing: chatbots are just one way to use Large Language Models (LLMs).

Unfortunately, I see many articles or presentations that just focus on demonstrating LLMs at work within the context of chatbots. I feel guilty of showing the traditional chat interfaces too. But there’s so much more to it!

Retrieval Augmented Generation (RAG) is a great way to expand the knowledge of Large Language Models to let them know about your own data and documents. With RAG, LLMs can ground their answers on the information your provide, which reduces the chances of hallucinations.

Implementing RAG is fairly trivial with a framework like LangChain4j. However, the results may not be on-par with your quality expectations. Often, you’ll need to further tweak different aspects of the RAG pipeline, like the document preparation phase (in particular docs chunking), or the retrieval phase to find the best information in your vector database.

My colleague Jaana Dogan was pointing at the Anthropic’s MCP (Model Context Protocol) documentation pages which were describing how to build MCP servers and clients. The interesting twist was about preparing the documentation in order to have Claude assist you in building those MCP servers & clients, rather than clearly documenting how to do so.

MCP tutorials are great. There are no tutorials really.

"Copy these resources to Claude, and start asking some questions like..." pic.twitter.com/GG50DMWNLW

Read more...

In my previous article, I detailed how to build an AI-powered short story generation agent using Java, LangChain4j, Gemini, and Imagen 3, deployed on Cloud Run jobs.

This approach involved writing explicit Java code to orchestrate the entire workflow, defining each step programmatically. This follow-up article explores an alternative, declarative approach using Google Cloud Workflows.

I’ve written extensively on Workflows in the past, so for those AI agents that exhibit a very explicit plan and orchestration, I believe Workflows is also a great approach for such declarative AI agents.

This project demonstrates how to build a fully automated short story generator using Java, LangChain4j, Google Cloud’s Gemini and Imagen 3 models, and a serverless deployment on Cloud Run.

Every night at midnight UTC, a new story is created, complete with AI-generated illustrations, and published via Firebase Hosting. So if you want to read a new story every day, head over to:

→ short-ai-story.web.app ←

The code of this agent is available on Github. So don’t hesitate to check out the code:

First article of the year, so let me start by wishing you all, my dear readers, a very happy new year! And what is the subject of this new piece of content? For a while, I’ve been interested in analyzing trends and topics in social media streams. I recently joined Bluesky (you can follow me at @glaforge.dev), and contrarily to X, it’s possible to access its Firehose (the stream of all the messages sent by its users) pretty easily, and even for free. So let’s see what we can learn from the firehose!

Yesterday, Google released yet another cool Gemini model update, with Gemini 2.0 Flash thinking mode. Integrating natively and transparently some chain of thought techniques, the model is able to take some more thinking time, and automatically decomposes a complex task into smaller steps, and explores various paths in its thinking process. Thanks to this approach, Gemini 2.0 Flash is able to solve more complex problems than Gemini 1.5 Pro or the recent Gemini 2.0 Flash experiment.

Hot on the heels of the announcement of Gemini 2.0, I played with the new experimental model both from within Google AI Studio, and with LangChain4j.

Google released Gemini 2.0 Flash, with new modalities, including interleaving images, audio, text, video, both in input and output. Even a live bidirectional speech-to-speech mode, which is really exciting!

When experimenting with AI Studio, what attracted my attention was AI Studio’s new starter apps section. There are 3 examples (including links to Github projects showing how they were implemented):