Module 1 of LLM Zoomcamp is done! 🎉

I turned my original RAG pipeline into an Agent!

I spent these last few days diving deep into Agentic RAG. It's been fascinating to build it step by step. Every time I ask the LLM to learn about something new, I see how it naturally figures out which tools to use, when to search, and how many times to gather info before giving me a solid answer.

What exactly is Agentic RAG?
It’s like giving the AI a brain that can actually act. Instead of just retrieving from a fixed knowledge base, the model decides whether it needs external tools first, gathers what it needs, and then answers. It’s pretty interesting to understand how it actually works behind the scenes!

Why does this matter?
A few days ago I asked for a detailed guide on using the OpenAI Python library with the chat.completion API. The Local LLM called web search multiple times until it had enough context and built something useful from those pieces. Now that I am building these systems, I can finally understand why it does what it does.

💡 Insights from this week:
- Building a static pipeline is a great start, but to make something truly flexible, you need function or tool calling. It lets the LLM look at the question first and decide whether it needs to search a knowledge base before answering.
- I used to think "chunking" was just about breaking up text. Turns out it can reduce token input by 3x! 🤯
- You have to learn how to walk before you run. Starting small, understanding each component manually, and seeing how the pieces fit together… it felt slow at first but worth it. Now I’m able to accelerate with agent frameworks like toyaikit, LangChain, PydanticAI, or OpenAI Agents.
- There is definitely a learning curve with the API syntax. Between the new response API and chat completions, tool responses are structured differently and you have to adjust your code accordingly. Frustrating at times, but also a great way to learn!

Quick takeaway:
It is best to start simple, then add complexity only when needed. Sometimes an agent can burn tokens unnecessarily, so only add that layer if your problem really needs it!

Had a lot of fun with this module and I’m already curious about what’s next. If you’re interested in learning along, this is the full free course Alexey at the Data Talks Club: https://github.com/DataTalksClub/llm-zoomcamp/

Anyone else tinkering with LLM agents lately? What kind of projects are you exploring or trying out? Would love to hear where your journey is heading!

#ai #localai #llm #mastodon #fediverse #buildinpublic #linux #github #aiengineering #DataEngineering

A tip for users of our experimental instance mementomori.social:

You can enable the algorithmic "For you" feed from the top right corner of the Home feed. When you turn it on, it will instantly switch to a ranked feed that prioritizes posts based on who you follow, who you comment on, and what you like or boost. If you also toggle on "Include posts from people you don't follow," you'll see trending content across the Fediverse based on your interests.

Unlike Trending tab, For you -feed is personal. The logic is similar to pre-Musk Twitter but user-focused and not corporate-owned.

The Trending tab remains unchanged, and /explore shows posts from roughly the past 6-12 hours, with a strong focus on the most recent ones. A post that gets massive engagement might stick around a bit longer, but most cycle out within half a day. Your own For you page changes every time you click the menu, refresh the page, or return.

This is ongoing, experimental work. The PR draft and source code are available here: https://github.com/mementomori-social/mastodon/pull/4

#MementoMoriSocial #Mastodon #MastoAdmin #OpenSource #BuildInPublic #Algorithm #AlgorithmicFeed #Algorithms

One feature I’m using every day: a unified inbox.

Replies from different platforms show up in one place, so I don’t have to jump between tabs to keep up with conversations.

Simple, but surprisingly useful.

#buildinpublic #saas #development #software #developer #marketing #ui

Today I wired the add-reading screen to a real Supabase insert in my senior-first health app.

The save invalidates four caches — dashboard, recent readings, trend chart, and history list. I centralised the history query key so the screen and the mutation share one source of truth rather than two separate constants.

Photo source is tracked for provenance, but image upload is a separate slice — a local URI doesn't resolve off-device.

Core write is live. #buildinpublic

The next version of CreatorCaps lets you start captioning straight from Photos.

No need to open the app first:
Select a video → Share → CreatorCaps.

A little less friction between recording & captioning.

Available on iPhone and iPad as a one-time purchase: https://apple.co/4sT5vRu

#IndieDev #iOS #VideoEditing #BuildInPublic

Picture this: two printouts slide across the table. Your git history. Your browser history. The highlighted dates line up: your commit lands 22 days after you opened the competitor's repo. Counsel asks, politely, what you were thinking.

That scene is why I keep a prior-art ledger. Reading their code is the protective move. Avoiding it just moves the collision somewhere more expensive.

https://praveenlavu.com/dispatch/prior-art-discipline

#buildinpublic #ip #opensourcelaw #engineering

Just built a personal system to manage my ADHD brain — arOS. It’s not perfect, but it helps me track tasks, debug AI fallbacks, and keep my self-hosted setup running.

#BuildInPublic #Laravel #PHP #SoloDev