Retrieval Augmented Generation (RAG) is a pattern to let you prompt a large language model (LLM) about your own data, via in-context learning by providing extracts of documents found in a vector database (or potentially other sources too).

Implementing RAG isn’t very complicated, but the results you get are not necessarily up to your expectations. In the presentations below, I explore various advanced techniques to improve the quality of the responses returned by your RAG system:

It’s always interesting to know how the sausage is made, don’t you think? That’s why, a while ago, I looked at embedding model tokenization, and I implemented a little visualization to see the tokens in a colorful manner. Yet, I was still curious to see how Gemini would tokenize text…

Both LangChain4j Gemini modules (from Vertex AI and from Google AI Labs) can count the tokens included in a piece of text. However, both do so by calling a REST API endpoint method called countTokens. This is not ideal, as it requires a network hop to get the token counts, thus adding undesired extra latency. Wouldn’t it be nicer if we could count tokens locally instead?

When integrating an LLM into your applicaton to extend it and make it smarter, it’s important to be aware of the pitfalls and best practices you need to follow to avoid some common problems and integrate them successfully. This article will guide you through some key best practices that I’ve come across.

Understanding the Challenges of Implementing LLMs in Real-World Applications

One of the first challenges is that LLMs are constantly being improved. This means that the model you start using could change under the hood, and suddenly your application doesn’t work as it did before. Your prompts might need adjustments to work with the newer version, or worse, they might even lead to unintended results!

The happy users of Gemini Advanced, the powerful AI web assistant powered by the Gemini model, can execute some Python code, thanks to a built-in Python interpreter. So, for math, logic, calculation questions, the assistant can let Gemini invent a Python script, and execute it, to let users get a more accurate answer to their queries.

But wearing my Apache Groovy hat on, I wondered if I could get Gemini to invoke some Groovy scripts as well, for advanced math questions!

The latest release of LangChain4j (version 0.31) added the capability of grounding large language models with results from web searches. There’s an integration with Google Custom Search Engine, and also Tavily.

The fact of grounding an LLM’s response with the results from a search engine allows the LLM to find relevant information about the query from web searches, which will likely include up-to-date information that the model won’t have seen during its training, past its cut-off date when the training ended.

As a follow-up to my talk on generative AI for Java developers, I’ve developed a new presentation that focuses more on the Gemini large multimodal model by Google.

In this talk, we cover the multimodality capabilities of the model, as it’s able to ingest code, PDF, audio, video, and is able to reason about them. Another specificity of Gemini is its huge context window of up to 1 million tokens! This opens interesting perspectives, especially in multimodal scenarios.

Lately, for my Generative AI powered Java apps, I’ve used the Gemini multimodal large language model from Google. But there’s also Gemma, its little sister model.

Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models. Gemma is available in two sizes: 2B and 7B. Its weights are freely available, and its small size means you can run it on your own, even on your laptop. So I was curious to give it a run with LangChain4j.

No need to be a Python developer to do Generative AI! If you’re a Java developer, you can take advantage of LangChain4j to implement some advanced LLM integrations in your Java applications. And if you’re interested in using Gemini, one of the best models available, I invite you to have a look at the following “codelab” that I worked on:

Codelab — Gemini for Java Developers using LangChain4j

In this workshop, you’ll find various examples covering the following use cases, in crescendo approach:

As I was working on tweaking the Vertex AI text embedding model in LangChain4j, I wanted to better understand how the textembedding-gecko model tokenizes the text, in particular when we implement the Retrieval Augmented Generation approach.

The various PaLM-based models offer a computeTokens endpoint, which returns a list of tokens (encoded in Base 64) and their respective IDs.

Note: At the time of this writing, there’s no equivalent endpoint for Gemini models.

So I decided to create a small application that lets users:

A promising feature of the Gemini large language model released recently by Google DeepMind, is the support for function calls. It’s a way to supplement the model, by letting it know an external functions or APIs can be called. So you’re not limited by the knowledge cut-off of the model: instead, in the flow of the conversation with the model, you can pass a list of functions the model will know are available to get the information it needs, to complete the generation of its answer.