David Mezzetti

One common take I've seen on the LiteLLM breach is this: pin your dependencies to a specific version. Yes, this would fix this specific issue but the vast majority of security risks are found and patched over time. You're likely more vulnerable if you don't upgrade vs if you do.

This 970K parameter model retains 98% of the performance of the original 110M medical embeddings model. huggingface.co/NeuML/biomedbe…

Want a solid medical literature re-ranker model? Then check this out! huggingface.co/NeuML/biomedbe…

Want a late interaction medical vector model? Then check this out! huggingface.co/NeuML/biomedbe…

TxtAI has embeddings databases, pipelines, agents and workflows. One little known but powerful feature of TxtAI is that it can export any of it's functionality as an OpenAI endpoint. Check out this example for now. github.com/neuml/txtai/bl…

It's always a good idea to rotate your keys on a recurring basis and/or have an expiration date to force this.

An automated publishing workflow is convenient but is it necessary? Thinking about the LiteLLM compromise yesterday, a manual step to push to PyPI certainly could have helped. Why? Because the API token wouldn't have been stored on a server.

FYI - Added this issue to txtai for awareness since LiteLLM is an optional dependency. github.com/neuml/txtai/is…

PSA: If you use LiteLLM and you installed a version this morning. Luckily the package is now quarantined but a good idea to rotate your tokens if you did use it. github.com/BerriAI/litell…

Top trending post on HN at the moment: news.ycombinator.com/item?id=475017…

A simple JEPA world model in 15M parameters! This is the exact kind of thinking we need not just brute forcing more GPUs at the problem. It's clear that efficiency is not good for everyone though. arxiv.org/abs/2603.19312…

💬 Talk is much cheaper than action. AI is amazing in many ways but it has trade offs like anything else.

Our PubMedBERT embeddings model is the most downloaded open model for medical vector embeddings. Over 1M downloads this month! huggingface.co/NeuML/pubmedbe…

If you want a Local AI Agent stack with dev tools, then check out this video on how to do that with TxtAI. youtube.com/watch?v=RDNaFX…

One of the most accessed notebooks on TxtAI's GitHub page covers the Semantic Graph. TxtAI can automatically build a graph of related nodes using it's vector similarity model. Learn more here. github.com/neuml/txtai/bl…

🚀 TxtAI v9.7 is out with the new Coding Agent Toolkit! Release Notes: github.com/neuml/txtai/re… github.com/neuml/txtai

Knowing what to build is much more important than the how. The best question an engineer asks is "Did you think of doing this instead of that?". So regardless of the technical stack, that part of the process is still an important human task.

Local AI might not always be the easiest way to do it but it's the only way you can guarantee control over your stack. Sending your every thought and request to AI Vendor APIs isn't the best idea for you long term.

Deep dive into the new TxtAI Agents Toolkit. It's worth the watch if you care about Local AI. youtube.com/watch?v=RDNaFX…

Vector search works well for non-exhaustive relevance-driven search. What if you need to run an exhaustive search? Straight keyword search with term expansion could be the best solution. There is no one-size-fits-all retrieval strategy.