Olivier Balais (overnetcity.bsky.social)

CTO @semji_fr, proud #casc co-founder, passionate fullstack web developer, I build cool things with #php #js #docker and an amazing team @SemjiTech !

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

Tested #Gemini 2.0 Flash to convert recordings+audio into specs for Windsurf — mind-blowing. Now hoping for native support :)

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

A few weeks ago, I was blown away by @windsurf_ai IDE! Next step: a multimodal approach (video+text+audio) so we can "show" code while explaining logic—like a screen record in a tech issue.

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

@dunglas Amazing! Thanks for the pointer, @dunglas! I checked the page you shared but couldn’t find any mention of the BoltDB you referenced. Is there any other documentation or guidance you could point me to?

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

Hey @dunglas, hope all’s well! I’m exploring https://t.co/SjdW8AQTpm for handling SSE auto-retries on network failure. Does Mercure have a built-in feature to store/buffer events for client auto-retry, or is it something we typically manage on the backend?

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

o3 just dropped and it’s a monumental leap in AI capabilities. From blazing code gen to near-human performance on the ARC AGI test (yes, that test), it changes what we thought AI could do. Exciting? Absolutely. Terrifying? A bit. Hardware is now the real bottleneck. Buckle up!

[Pinecone @pinecone]

Our new models — fully integrated alongside our database — bring best-in-class retrieval to your applications:

✔️ Our new sparse embedding model — pinecone-sparse-english-v0 — boosts performance for keyword-based queries, delivering up to 44% and on average 23% better NDCG@10 than BM25 on TREC.
✔️ Our new reranking model — pinecone-rerank-v0 — improves search accuracy by up to 60% and on average 9% over industry-leading models on the BEIR benchmark.
✔️ @cohere's latest model — cohere-rerank-v3.5 — balances performance and latency for a wide range of enterprise search applications.

Learn more by visiting our Model Gallery: https://t.co/AaUJqbB3WH

[Pinecone @pinecone]

First-of-its-kind Pinecone Knowledge Platform Powers Best-in-class Retrieval for Customers

💠 Industry-leading vector database capabilities combined with proprietary AI models help developers build up to 48% more accurate AI applications: faster & easier

https://t.co/Pz9GkmUUe8

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

Bye bye Twitter. Let’s meet on Bsky: https://t.co/6ciohU3wz2.

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

I’ve been working and building products with generative AI for over 4 years now. It’s hard to impress me at this point. But damn, @windsurf_ai, your IDE is absolutely next-level! 🔥

[Visual Studio Code @code]

New in Copilot Chat... enhanced links for any workspace symbols that Copilot mentions 🔗

These links appear in responses as little pills, letting you jump directly to definitions for better understanding.

[goosewin @Goosewin]

build failed

[Jerry Liu @jerryjliu0]

Pretty excited about this new RAG technique I cooked up 🧑‍🍳

A top issue with RAG chunking is it splits the document into fragmented pieces, causing top-k retrieval to return partial context. Also most documents have multiple hierarchies of sections: top-level sections, sub-sections, etc.

This is also why lots of people are interested in exploring the idea of knowledge graphs - pulling in "links" to related pages to expand retrieved context.

This notebook lets you retrieve contiguous chunks without having to spend a lot of time tuning the chunking algorithm, thanks to GraphRAG-esque metadata tagging + retrieval. Tag chunks with sections, and use the section ID to expand the retrieved set.

Check it out

https://t.co/mIolxuMT12

[Rohan Paul @rohanpaul_ai]

Text chunking now matches human reading patterns by detecting natural breaks in information flow.

Meta-chunking, proposed in this paper, uses probability patterns to find natural segment boundaries in documents, just like humans do

Original Problem 🎯:

Text chunking in Retrieval-Augmented Generation (RAG) systems often fails to maintain logical coherence between segments, leading to incomplete or fragmented information retrieval. Current methods rely on fixed-length splits or basic semantic similarity, missing crucial logical connections between sentences.

-----

Solution in this Paper ⚡:

• Meta-Chunking: A novel segmentation technique operating between sentence and paragraph levels

• Two key strategies:
- Margin Sampling: Uses LLMs for binary classification to determine segment boundaries based on probability differences
- Perplexity (PPL) Chunking: Analyzes perplexity distribution to identify natural text boundaries

• Dynamic combination approach to balance fine and coarse-grained segmentation

• KV caching mechanism for handling longer texts efficiently

-----

Key Insights 💡:

• Smaller models (1.5B parameters) can effectively perform chunking tasks

• PPL distribution characteristics guide optimal threshold selection

• Dynamic chunk sizing preserves logical integrity better than fixed-length approaches

• Re-ranking performance improves significantly with Meta-Chunking

-----

Results 📊:

• Outperforms similarity chunking by 1.32 on 2WikiMultihopQA while using only 45.8% processing time

• PPL Chunking with Qwen2-1.5B achieves 0.3760 BLEU-1 score on single-hop queries

• Maintains consistent performance across both Chinese and English datasets

• Shows 3.59% improvement in Hits@8 metric when combined with PPLRerank

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

Peur sur les pistes cyclables : gros pneus et coups de sonnettes, la menace de la « SUVisation du vélo » – via @lemondefr https://t.co/jZKXMoKylC

[Marko Denic @denicmarko]

It's true.

[Yann LeCun @ylecun]

I don't want to say "I told you so", but I told you so.
Repeatedly.
I was heavily criticized for it. Mocked, even.

[Rohan Paul @rohanpaul_ai]

This project tries to implement a real-time replication of OpenAI’s groundbreaking O1 model.

Exploring advanced reasoning capabilities and a specific "journey learning" mechanisms for AI.

They propose a new approach: “journey learning”. This paradigm goes beyond the traditional focus on specific tasks and emphasizes continuous progress through learning, reflection, and adaptation.

[Olivier Balais (overnetcity.bsky.social) @overnetcity]

🤯

[Brad Costanzo @BradCostanzo]

Wow! @HeyGen_Official just released today ability to have an AI avatar join a Zoom meeting and interact.
I invited one of their AI avatars into a Zoom room and recorded this clip. Time to build my own now

[Brad Costanzo @BradCostanzo]

Wow! @HeyGen_Official just released today ability to have an AI avatar join a Zoom meeting and interact.
I invited one of their AI avatars into a Zoom room and recorded this clip. Time to build my own now