French hatred of us English is pure, decent, noble loathing!! How dare Russia try to muscle in on a relationship so special it could never understand??
Your documents are not general-purpose,
Your retriever should be good beyond these
@tomaarsen recently fine-tuned mLateOn-unsupervised for medical retrieval using MIRIAD, a benchmark with 1,000 queries and 200,000 passages.
Out of the box, mLateOn reached 0.8520 NDCG@10.
After a short fine-tuning run on a single RTX 3090, mLateOn-medical reached 0.9139, outperforming every general-purpose retriever tested. Qwen3-Embedding-4B, the strongest dense model in the evaluation, reached only 0.7817, despite having roughly 33 times more active parameters.
The difference comes down to how information is represented. Dense models compress an entire document into one vector. Late-interaction models preserve token-level representations, allowing them to capture signals that single-vector models may average away.
This matters even more for long, specialized documents. In this evaluation, passages averaged 941 tokens.
The conclusion is simple: strong retrieval is not only about building larger models. It is about choosing an architecture that can understand your documents and adapt to your domain.
→ Explore Tom Aarsen’s complete fine-tuning guide :
https://t.co/Z1mlZ4DgVA
→ Put advanced retrieval to work with LightOn Console
https://t.co/kY55OAXWIq
Need a model that, out-of-the-box, (almost) tops a specialist benchmark? mLateOn B)
Need to actually top that benchmark? Fine-tune it on an RTX 3090.
Late interaction just can't stop winning!
AI agents scale best when their economics scale with them.
@LightOnIO makes this possible by delivering the right context to your models before generation.
Better signal. Leaner prompts. Cleaner reasoning.
As agents scale, token consumption compounds fast.
In the past year alone:
• Some agentic tasks used up to 650× more tokens
• Frontier model pricing rose 4× in just eight months
At scale, every unnecessary token becomes a recurring cost.
With retrieval-first AI, agents can use fewer tokens, reduce inference costs, and operate more efficiently.
Estimate your potential savings with the LightOn Agent Economics Simulator.
Run the numbers → https://t.co/Yau404nxVn
The free lunch just got bigger.
After introducing hierarchical pooling, enabling 2× lossless compression for ColBERT state-of-the-art embedding models powering LightOn, we pushed the idea even further.
By training directly for compressed representations, hierarchical pooling now reaches 5× near-lossless compression without degrading full-model performance.
Same retrieval pipeline.
Same retrieval quality.
Less than a fifth of the storage footprint.
Read Antoine's thread and dive into the technical details ↓
https://t.co/GYpBcl1tcu
Every great agent starts with the right context.
LightOn Console plugs into your existing stack, connects to your LLM provider and your data sources, and gives every agent the exact signal it needs to solve the most complex tasks.
No new stack. No migration. Just better retrieval, lower token consumption, and more room for reasoning.
Build the agent you've always imagined.
👉🏻 Explore the integrations → https://t.co/djsvYo7jyJ
One API. Three endpoints. Every block a SOTA model.
LightOn Console is document intelligence you can build on. Parse, search, and reason over your documents through one API, called by you or by your agents.
On OCR and retrieval benchmarks, LightOn models go head to head with the world's largest AI labs, and win where it counts.
Benchmarks are where it starts. Production is where it's decided.
✅ Heavy parsing loads
🔍 Retrieval that scales
🔐 Access rights preserved
🌍 Deployment where your data is allowed to live
That is what Console is built for.
State-of-the-art OCR and retrieval, with auditable citations, sovereign deployment, and built from the ground up for agents and the people who direct them.
A strong pipeline starts with strong models. A pipeline you trust is one you can test and audit in production.
Console gives you both.
📊 Explore the benchmarks → https://t.co/BTclbsI9U6
⚙️ Test LightOn Console → https://t.co/JE8h5QZyPZ
I'm glad you asked!
https://t.co/5NQuTtiBKy - Apache 2.0
https://t.co/T0n6V4Dn9S - Apache 2.0
https://t.co/vAf0oNZEwO - Apache 2.0
https://t.co/y2FVSOvmT2 - Apache 2.0
https://t.co/m5MElIXLwt - Apache 2.0
https://t.co/n3bvxNx1Br Apache 2.0
https://t.co/SPXHsuizqQ - Apache 2.0
https://t.co/gM3wI7MJF9 - Apache 2.0
https://t.co/HsBKXKFufO - Apache 2.0
https://t.co/MTdJbtb7NJ - Apache 2.0
https://t.co/kbEn0WmOph - CC-BY-NC 4.0 (but this because of the data)
https://t.co/egnC6HKgzK - Apache 2.0 (lifting the previous CC BY NC license + adding 10% improvement
https://t.co/Cw3FuO55YR - Apache 2.0
https://t.co/s9fyUCHDUg - Apache 2.0
https://t.co/lZXjheCq0s - MIT
https://t.co/eVNl8J3yI5 - Apache 2.0
https://t.co/3OOQsL4si2 - Apache 2.0
The good part of building the frontier in the open is that you do not have to rely on anything with closed license. We even make the effort for everyone to be able to use them and even build their own by giving code and data everytime
The bad part is that people will never be happy and will always want more. It feels like they never want to actually build anything from themselves
And you, @agenticrypto, what's your licenses?
🎬 Les cas d’usage impossibles – Saison 1, épisode 08 : Ingénierie aéronautique
L’audit qui a évité l’immobilisation d’une flotte d’avions.
📄 Une consigne de navigabilité EASA rédigée en anglais.
✈️ Des servovalves AV-3000 à identifier sur des appareils en service, opérateur par opérateur.
🔎 Des preuves d’incorporation dispersées dans des certificats EASA Form 1, parfois rangés dans le dossier d’un autre appareil.
Le risque n’est pas seulement de manquer une pièce défectueuse.
Il est de déclarer un avion conforme sans preuve, ou d’immobiliser inutilement un appareil déjà traité.
Un RAG classique retrouve la consigne.
LightOn Console produit l’état réel du parc, unité par unité : quels appareils sont concernés, lesquels sont déjà mis en conformité, lesquels restent à traiter, et sur quel document repose chaque réponse.
Dans l’aéronautique, l’audit qualité ne sert pas seulement à retrouver des documents. Il peut décider quels avions peuvent voler demain.
📰 Lire l’analyse complète : https://t.co/fybcKxlton
💻 Testez le scénario sur LightOn Console : https://t.co/Kcs9s21NLy
Meet LightOn Facets.
When faceted search meets agentic retrieval,it brings the precision every AI needs. More signal, less noise, at scale.
Agents don't just run your filters. They discover the right facets from a plain query, then scope the search themselves. What reaches the model is already signal.
Now live in LightOn Console : https://t.co/pgCp2D9p9J
So did something fun!
As I am preparing a lecture on late interaction, reading papers on efficient retrieval, I did what any serious researcher would do nowadays:
downloaded the papers, gave them to an agent, and asked it to build a beautiful frontend to explore them!
Fewer Tokens, Better Answers: Give Your Agent Search, Not Raw Files
Even with just 295 pages, agents benefit from search.
We ran the same Claude agent on the same corpus two ways:
📂 Using its native tools to explore the corpus, inspect documents, and decide what to read
🔍 Using LightOn API for Search
On a 5,000-page knowledge base, LightOn delivered:
✅ 36% fewer tokens
✅ Better answers in 75% of blind evaluations
✅ Flat search costs as the corpus grew 17×
The interesting part isn't that long context windows eventually hit their limits.
It's that search already improves efficiency and answer quality long before context becomes the bottleneck.
A larger context window gives your agent more room, but it doesn't magically make it better at retrieval.
Full breakdown 👇🏻
https://t.co/DnW5Af3dZr
Start saving tokens now with : https://t.co/R3qb8NStdT
💬 « Nous avons choisi @LightOnIO parce que c'était l'équipe fiable, à l'écoute et réactive avec laquelle nous voulions construire sur le long terme. »
Vincent Kerbiquet, Directeur d'Infocom'94
@infocom94 , opérateur public numérique mutualisant les services de 26 collectivités en Île-de-France, a choisi LightOn pour déployer une IA conversationnelle souveraine au service de ses adhérents.
La confiance n'est pas un supplément. C'est le point de départ.
👉🏻 Plus d'informations : https://t.co/0kwgYqM90T
🎬 Les cas d’usage impossibles – Saison 1, épisode 07 : RH
La due diligence RH… qui pouvait sauver une cession.
📄 Des clauses de non-concurrence dispersées entre contrats, modèles pays, avenants scannés et textes de droit du travail.
🌍 Des clauses françaises, allemandes, britanniques, américaines et marocaines à interpréter différemment.
🔎 Une clause déjà levée, absente des récapitulatifs, mais décisive pour l’inventaire.
Le risque n’est pas de manquer une clause. Il est de croire qu’une clause existe encore, ou qu’elle engage, alors qu’un tribunal pourrait l’écarter.
Un RAG classique retrouve les bons contrats.
LightOn Console produit une analyse sourcée, juridiction par juridiction, exploitable dans une data room avant que le conseil de l’acquéreur ne trouve la faille en premier.
Dans une opération de cession, l’audit RH ne sert pas seulement à documenter le risque. Il peut sécuriser la transaction.
📰 Lire l’analyse complète : https://t.co/6t86FLc5of
💻 Testez le scénario sur LightOn Console : https://t.co/Kcs9s21NLy
7 million pages of legal documents. One OCR model.
That's what it took to build the LOCUS dataset, one of the largest publicly available collections of U.S. local laws.
To process millions of legal documents published across thousands of cities and counties, @DenisPeskoff , @barrowjoseph , Christopher Vu, and Diag Davenport relied on LightOnOCR-2-1B to convert heterogeneous documents into a consistent, structured corpus, combined with ModernBERT for downstream processing and analysis.
Why LightOnOCR-2?
⚡️ Up to 3× faster than larger OCR models
🎯 State-of-the-art performance
📄One single model, from document to markdown
📦 Just 1B parameters, making large-scale processing practical
When you're processing millions of documents, accuracy isn't enough. You need speed, efficiency, and reliability at scale.
🔗 Dataset: https://t.co/jDPlTxHVAo
🔗 Try LightOnOCR-2 in the LightOn Console: https://t.co/R3qb8NStdT
Thanks to @LightOnIO's technology (LightOnOCR, ModernBERT), the local codes of all 9,239 US cities and counties 🇺🇸, 7M pages OCR'd into 2.2M analyzable sections (~80GB raw), are now machine-readable and analyzable at national scale.
Outstanding work by @DenisPeskoff, @barrowjoseph & team (@UCBerkeley).
For more on using LightOn's technology at scale try it for free at: https://t.co/b9fz6WvhTx