Articles and sources curated by Yury
RSS FeedPostgres creator and DBOS co-founder Dr. Mike Stonebraker explains why workflow architectures are preferred over event-driven and why AI frameworks are converging on a workflow-style model and supporting durable execution.
How to implement a virtual filesystem over persistent data storage, such as an Elasticsearch database, to give AI agents access to shell commands, like grep and cat.
Build knowledge graphs from unstructured text using Claude for entity extraction, relation mining, deduplication, and multi-hop graph querying.
Learn what GRPO is and how to fine-tune LFM2.5-1.2B-Instruct with GRPO and Unsloth on OCR receipt extraction for JSON.
In the first piece, I argued that a real Company Brain needs three kinds of memory: factual memory, interaction memory, and action memory. F...
Almost everything important in a company happens in meetings, messages, or emails.That sounds exaggerated until you start listing the actual...
LlamaIndex is a simple, flexible framework for building knowledge assistants using LLMs connected to your enterprise data.
Ship RAG with Weave CLI: Opik traces, an LLM-judge eval harness, and config-driven swaps across 11 vector databases, chunking, and agents.
One of the hardest parts of any organization is institutional friction. Conversations lose context. Meetings create ambiguous follow-ups. Pe...
Agentic applications—AI systems empowered to take autonomous actions by calling external tools—are the current rage in software development. They promise efficiency, convenience, and reduced human intervention. Giving autonomous agents access to tools with potentially irreversible side effects,
The Silicon Valley narrative centers on entrepreneurial protagonists who are poised one predestined step away from changing the world. A decade ago they were heroes, and more recently they’ve become villains, but either way they are absolutely the protagonists. Working within the industry, I’ve
I’ve come to believe that most organizational design questions can be answered by recursively applying a framework for sizing teams. Over the past year I’ve refined my approach to team sizing into a bit of a framework, and even changed my mind on several aspects, especially the viability of smal
A deep dive into LiteParse's grid projection algorithm — how it extracts text from PDFs while preserving tables, columns, and alignment. Open source.
There are only 6 fundamental ways to evaluate a RAG system. Here's how to do it.
Use cases, techniques, alignment, finetuning, and critiques against LLM-evaluators.
Good evaluations help teams ship AI agents more confidently. Without them, it’s easy to get stuck in reactive loops—catching issues only in production, where fixing one failure creates others. Evals make problems and behavioral changes visible before they affect users, and their value compounds
In search we talk about lexical embedding retrieval But we miss another retrieval philosophy metadata And with LLMs its never...
There are many ways to slice a Postgres database for multi-tenant applications. Let's look at the three most common approaches and the trade-offs.
Giorgos Nikolaou <sup>1,5,</sup> Tommaso Mencattini <sup>1,2,∗</sup> Donato Crisostomi <sup>2</sup> Andrea Santilli <sup>2</sup> Yannis Panagakis <sup>4,5</sup> Emanuele Rodolà <sup>2,3</sup>
The best agents fail at enterprise documents: not because they can't reason, but because they can't read. We’re announcing Document Intelligence: powerful state-of-the-art research turning your unstructured enterprise documents into agent-ready data, at scale.
Our AI tooling survey finds concerns about mounting AI costs, more engineers hitting usage limits, and AI tools having uneven effects upon different types of engineers
By Ryan Lopopolo, Member of the Technical Staff
Explore techniques to scale late interaction models like ColPali for large-scale vector search in Elasticsearch.
Learn about ColPali and explore how to use it to search through complex documents in Elasticsearch, including tables, figures, multiple columns & more.
Inside Meta’s Home Grown AI Analytics Agent From Hack to Company-Wide Tool The hypothesis was simple: can an AI agent perform routine data analysis tasks autonomously? Data scientists tend to get …
A practical, reproducible approach to unsupervised document clustering with Elasticsearch and Jina embeddings.
Zhanghao Hu Qinglin Zhu Hanqi Yan Yulan He Lin Gui
Letta Filesystem scores 74.0% of the LoCoMo benchmark by simply storing conversational histories in a file, beating out specialized memory tool libraries.
Andy Nguyen <sup>1</sup>, Danh Doan <sup>1</sup>, Hoang Pham <sup>1</sup>, Bao Ha <sup>1</sup>, Dat Pham <sup>1</sup>, Linh Nguyen <sup>1</sup>,
TL;DRForgetting and false recall are mathematically, provably, inevitable for any memory system that organises information by meaning. Not j...
Late interaction allow for semantically rich interactions that enable a precise retrieval process across different modalities of unstructured data, including text and images.
TL;DRMemory systems for LLMs forget EXACTLY like humans, reproducing the exact numbers from some of the most replicated experiments in clini...
Most agent runtimes reduce complex documents to garbled text, degrading every downstream task. LlamaParse and LiteParse are installable agent skills that deliver structured, layout-aware, multimodal document parsing — locally or via cloud — so agents can finally reason over real enterprise docum
We're introducing Context Repositories, a rebuild of how memory works in Letta Code based on programmatic context management and git-based versioning.
Two hidden LLM failure modes that break agentic document ingestion pipelines in production: repetition loops exhaust resources while recitation filters block legitimate tasks. Learn how to detect, handle, and mitigate both across OpenAI, Anthropic, and Gemini APIs.
Best practices for writing database retrieval tools for context engineering. Learn how to design and evaluate agent tools for interacting with Elasticsearch data.
The highest-scoring AI memory system ever benchmarked. And it's free. - Commits · milla-jovovich/mempalace
We replaced expensive sandboxes with ChromaFs, a virtual filesystem over Chroma, to give our docs AI assistant the ability to explore documentation like a developer would.
Three patterns for building on the Claude Platform that keep pace with Claude's evolving intelligence while balancing latency and cost.
How coding agents use tools, memory, and repo context to make LLMs work better in practice
Userspace filesystem:
One of the hardest parts of building an agent harness is constructing its action space.Claude acts through Tool Calling, but there are a num...
Murphy Campbell has an AI imposter on Spotify and troll claiming copyright of her performances of public domain ballads on YouTube.