Breaktime Tech Talks

Ep79: Reducing Agent Token Costs + RAG Beyond Semantic Search

In this episode, I sit down with Roie Schwaber-Cohen, a software engineer and developer advocate at Pinecone, to talk about smarter ways to build with AI — without burning through tokens or your patience!
What we cover:
Why agentic AI systems burn so many tokens (and ways to combat it)
How Pinecone's Nexus pre-explores retrieval paths so agents don't have to discover them at runtime, cutting latency and token usage
The problem with naive RAG ("Franken answers") and why domain-level separation of your documents matters
How Pinecone Marketplace lets non-developers connect structured and unstructured data sources to build production-ready AI apps
Why semantic similarity isn't the same as correctness, and how document introspection helps agents ask better questions
Links & Resources:
Pinecone
Pinecone Marketplace (recently announced)
Pinecone Nexus
Roie on LinkedIn

Ep78: AI-First Java Book + Embabel for Enterprise Agents

I'm back after a couple of weeks of hiatus with a packed update. From a major book deadline to enterprise graph hackathons, summer is anything but slow.
AI-First Java Book. First six chapters officially submitted. They cover concept progression, real-world problem solving, and a developer's career journey
Customer Graph Hackathons. First hands-on event with Neo4j at an enterprise customer site, and another one coming next week
GraphRAG Fundamentals Training. Rescheduled on O'Reilly Learning Platform; available to sign up
Neo4j CLI. New tool for interacting with Neo4j from the command line, with agent skill support for coding tools like Claude
Agent Instruction Protocol. Open-source repo that turns skill specs into YAML-based execution graphs modeled as process flows
Building Agents in Java with Embabel. Dan Vega's walkthrough of this Java AI framework. It covers actions, plans, goals, and reusable components for enterprise agents
Lots of exciting things in motion — grab the links in the show notes and happy coding!

Ep77: Spring AI Memory Bug + LLMs, Skills, & Your Dev Career

A packed week of travel, debugging, writing, and reading. I share what I learned and ran into this week as a developer.
Highlights:
🗣️ Spoke at the JavaMUG (Java Metroplex User Group) meetup in Dallas — highly recommended for Dallas-area developers!
🐛 Discovered a bug in Spring AI's Neo4j chat memory integration with the latest Spring Boot — workaround is to use in-memory storage for now.
📖 Working on a Java book chapter covering OOP concepts (abstraction, encapsulation, polymorphism) with a focus on practical, use-case-driven learning
Event updates:
📅 Session from the AI Agents conference is now on YouTube
GraphRAG Fundamentals O'Reilly course rescheduled to June in APAC timezone
NODES CFP open until June 15th
📝 Two recommended reads:
"MCP, Skills, and Agents" by David Cramer
"LLMs and Your Career" by Phil Eaton

Ep76: Batching in Native Cypher + Avoiding Injection Vulnerabilities

A quieter week, but still full of forward motion — from clearing the Neo4j developer blog backlog and making progress on the upcoming Java book, to lining up upcoming speaking events. Plus, two developer-focused content pieces I hope you enjoy as much as I did.
This Week's Updates:
Rescheduling the GraphRAG Fundamentals training (likely June, APAC-friendly time zones)
Cleared a backlog of community submissions for the Neo4j Developer Blog — open to anyone with a Medium account
Cypher/SQL injection understanding from Neo4j Definitive Guide book
Writing progress on a new AI-first Java learning book, drawing on literary aspirations and music pedagogy principles for a fresh teaching approach
New speaking opportunities coming in May and June — stay tuned
Content Pieces:
"Batching Like a Pro" by Gemma Lamont — A detailed feature comparison of apoc.periodic.iterate vs. native Cypher CALL IN TRANSACTIONS, covering memory tracking, error handling, query planning, concurrency, entity rebinding, and retry strategies. The gap has largely closed — native Cypher is nearly on par.
NODES AI videos now on YouTube — All session recordings from Nodes AI are publicly available. Link in the show notes.

Ep75: Versioning AI Models + Define Schema for Better Knowledge Graphs

This week, I prepped for upcoming events, tweaked and strategized some existing processes, and found more data on how defining a schema can produce better knowledge graph construction.
Highlights:
Prepped for two upcoming events: a Graph RAG Fundamentals training on O'Reilly Learning Platform and a session at a virtual AI Agents conference.
Updating repositories for the workshop surfaced a chain-reaction lesson: upgrading frameworks leads to data changes, which require config updates, which require prompt rewrites.
Key takeaway — don't pin your apps to latest for AI models, just as you wouldn't for Docker image tags. Tie to a specific version so updates don't cascade unexpectedly.
Also revisited my tech blogging workflow and built a template script to eliminate boilerplate setup, shaving time off the writing process without sacrificing the actual content creation.
New blog post on agents, tools, and MCP published in the process!
On the Neo4j side, I touched on the Neo4j Educator Program and how learning patterns among new developers are shifting — happy to accept feedback from educators teaching graphs.
This week's article is Hands-on KG Relation Resolution by Mike Dillinger. It examines knowledge graph construction and why defining a narrowed schema produces cleaner, more understandable graphs. Without boundaries, LLMs and NLP processes generate overly granular, spaghetti-like structures.

Ep74: Building without a Blueprint + Scaling Graphs with Infinigraph

In this episode, I reflect on career growth in tech after speaking with a group of students, along with a few technical topics I explored this week — from Cypher optimization to scaling graph databases.
💡 Highlights
Career growth isn’t linear — most skills come from experimenting, saying yes to opportunities, and building over time rather than formal training
Project ideas come from doing — exploring tools, creating content, and solving real problems often reveal what to learn next
Cypher optimization (GraphAcademy) — hands-on practice with EXPLAIN, PROFILE, and query tuning reinforces key performance concepts
Neo4j Infinigraph — a new approach to scaling graphs by separating graph structure from large property data, improving performance and scalability (plus, NODES AI website and NODES AI YouTube playlist)
A reminder that progress comes from building, exploring, and iterating — not waiting for the perfect plan.

Ep73: Cypher Query Tuning + Does Language Still Matter in the Age of AI?

Back from a short holiday, I caught up on a few things this week — including the inevitable yak shaving.
Highlights:
📖 Book Progress: wrapped up another chapter draft. I've been finding that blocking larger chunks of dedicated time makes a real difference for focus and momentum, although getting started is still the hardest part.
🎓 Neo4j Educator Program: spent time refreshing slide decks, resource links, emails, and tutorials for the program. Still more to do, but happy with the progress. (Also worth knowing: the Neo4j Startup Program now offers Aura cloud database credits depending on your stage.)
🐃 Yak Shave of the Week: had to clear disk space on my laptop just to run required software updates — for the second time recently. Frustrating but necessary.
Content:
📚 Learning: Neo4j: The Definitive Guide – Ch. 5 (Query Analysis & Tuning) A fantastic chapter covering how the Cypher Query Planner works (Pipeline, Slotted, and Parallel), plus deep dives into EXPLAIN and PROFILE for query optimization. Exactly the kind of under-the-hood content I've been looking for.
💬 Reading: "Does Language Still Matter in the Age of AI?" — David Parry A great read on why structured, verbose languages actually perform better in AI code generation — and are easier to review. Language expertise is still very much worth developing.

Ep72: Live Coding Fails and Fixes + Why LLMs Lose the Plot

My recap of virtual presentations, live streams, and workshop support — sharing wins, lessons from a humbling live coding session, and a fascinating article on solving long-running LLM memory problems.
Highlights:
Delivered a virtual meetup for the San Francisco ACM on building knowledge graphs with the Neo4j GraphRAG Python package (code repository)
Helped as a TA during a Road to Nodes AI workshop covering MCP server integrations with Neo4j.
Attempted a live stream refactoring Postgres to Neo4j using OGM — ran into challenges that revealed documentation gaps and learning opportunities
Progress on the AI Java book with a productive working session
Shared a blog post by James Dunham: "Long Running LLM Conversations Need Working Memory, Not Just More Context" — which mirrors issues encountered in a prior RPG project where LLMs lost story continuity over time
Upcoming: Road to Nodes AI workshop on long-term memory & agentic workflows (free, virtual)
Upcoming: NODES AI virtual conference — April 15th (free)

Ep71: GraphRAG Learnings + Langchain4j Apps for Production

This week, I share hard-won lessons from building a GraphRAG application with Neo4j in Python, plus standout tips from Lize Raes's Devoxx Belgium talk on taking Langchain4j apps to production.
GraphRAG with Neo4j
Built a Python GraphRAG app using the Neo4j GraphRAG package — knowledge graph construction, retrievers (vector, graph, text-to-cypher), and agentic orchestration
Key lesson: don't let the LLM decide your entire data model. Providing node types, relationship types, and patterns as boundaries dramatically improves results
Expect iteration — retrieval testing will send you back to refine your KG construction
Github code: Neo4j GraphRAG Python package
Langchain4j for Production (Lize Raes, Devoxx Belgium)
Wrap RAG as an agent tool for multi-call retrieval instead of single-shot pipelines
Filter available tools programmatically by domain to keep agents focused
Wire sub-agents as @Tool for clean multi-agent orchestration
Use immediate responses to skip the LLM summarization hop — saves tokens and latency
13-step walkthrough for production-grade agentic systems
YouTube link: Level Up Your Langchain4j Apps for Production (Lize Raes, Devoxx Belgium 2025)

Ep70: Devnexus Conference Recap + No-LLM Agentic Workflow

Hear my recent experience at the Devnexus conference in Atlanta, where I delivered two sessions and connected with so many amazing people!
Devnexus session 1: "Agents, Tools, and MCP, Oh My! Next Level AI Concepts for Developers" — a redesigned solo talk breaking down AI building blocks (agents, tool calls, context management, memory, and MCP) so developers can mix and match components for their own stack.
Key takeaway: AI systems are much more than just the LLM — developers play a critical role in designing the surrounding architecture.
Devnexus session 2: "Supercharging Applications with Java, Graphs, and a Touch of AI" (code repo 1, code repo 2) — a joint session with Erin Schnabel building an LLM-powered role-playing game using Langchain4j, Quarkus, and Neo4j.
Multiple approaches: plain LLM chat, prompt engineering, and RAG with Neo4j as the vector/graph store, chunking documents while preserving structure via graph relationships.
Our "Three Cs" challenge: Continuity (maintaining storyline), Context (growing context window), and Creativity (keeping the LLM on track without going off the rails).
Splitting responsibilities between the LLM and a deterministic engine significantly improved results — a pattern developers should consider for complex AI apps.
App redesign with an agentic architecture: dice roll, narration, suggestion, checkpoint, and recap agents — with the last three running concurrently for better performance.
Markdown file (in one app) for agentic memory, enabling easy edits, rollbacks, and incremental indexing during live gameplay.
Content spotlight: "No Keys, No LLM — Building a Wikidata Definition API with Embabel" — an article showcasing an agentic Java application that uses zero LLM. Embabel (a Java agentic framework) handles planning and execution with structured inputs/outputs, no external or local model required.
Could the no-LLM agent pattern see broader adoption, or is it a niche experiment?
New episodes will now use platform-agnostic Podfollow links.
New blog post on jmhreif.com about Cypher AI procedures.