.txt

?? NEW PODCAST: ???????????????????? ???????????? ???????? ???????????????? with Cameron Pfiffer from .txt and Stefan Webb ?? Cameron shares his journey into the field, the mission of .txt, and the importance of constrained sampling in AI development. ?? Watch: https://lnkd.in/g4teUtiB ??? Listen: https://lnkd.in/g3tdUMp7 #Podcast #AIPodcast #Developers

Outlines v1.0 has a few interface changes coming up. One is the addition of the Application class, which combines a prompt templating function and an output type. Then, you call that application with the language model you want to use, as well as the arguments to the prompt function.

Looking forward to being at?NVIDIA #GTC2025 next week, with Alexis, returning to the Bay Area, and (re)connecting with entrepreneurs, VCs, business Leaders, and former colleagues. Great also to see there some of Elaia portfolio companies with Rémi Louf from .txt, Théau Peronnin from Alice & Bob and Dali Kilani from FlexAI! Thanks Serge Lemonde and Howard Wright for the opportunity! #computing #robotic #AI #semiconductor #photonics

Here's what happens if you disable the use of "R" and "r" in the thinking block of Qwen 7B R1. We asked if roses are red, and it invented the phrase "conclusion point" to use instead of "answer". It eventually got the right answe. We refer to this as "thought control". The thinking block that R1 and other reasoning models use can be guided using structured generation. You can force certain formats, remove or add words. Thought control works for anything you can express with a regular expression or context-free grammar. We'll be exploring this more in the future -- stay tuned.

Outlines just reached 11,000 stars. Thank you ??

Is that our advocacy team Will Kurt and Cameron Pfiffer hanging out with Andrew Ng? Wonder why.

What features do you want to see in Outlines?

My first blog post is out! ?? In this post, we'll explore two great libraries and you'll learn: - How to use the `outlines` library by .txt to make LLMs generate valid JSONs. - How to build AI applications using Haystack by deepset. - How to create custom components to use `outlines` within the `haystack` framework. Interested in using structured generation in your Haystack AI application? Check out `outlines-haystack`, link in the comments! ?? https://lnkd.in/dKtzMxxR

Structured Generation as the Foundation of Agentic Graph Systems ?? The intersection of structured generation (particularly the .txt approach) and agentic graph systems represents a critical evolution in AI system architecture. Structured generation emerges as a fundamental requirement—not just a beneficial feature—for effective knowledge graph construction, maintenance, and reasoning. The "New Rules for AI" manifesto introduces .txt as a tool designed to control LLM outputs through constraints, addressing what the authors identify as the "syntax problem" between human language understanding and computer syntax requirements. This capability appears to be foundational to the entire agentic graph system architecture that requires consistent, reliable outputs to maintain knowledge graph integrity. LLMs excel at understanding and generating human language but struggle to consistently produce outputs that conform to structured requirements of computer systems. This creates a "syntax disconnect." The .txt approach constrains LLM outputs at generation time, functioning as guard rails that ensure outputs follow precise formats required for knowledge graph construction and reasoning. Without this capability, the probabilistic nature of LLMs would conflict with the deterministic requirements of graph databases, creating an insurmountable barrier to reliable system integration. There is an evolutionary spectrum of structural control approaches, from direct Cypher generation (least structured) to RDF triples to JSON with schema validation (most structured). This progression demonstrates how increasing structural constraints leads to better knowledge graph quality, with the JSON schema validation approach producing "the richest and most consistent results." This suggests that the degree of structured generation directly correlates with the quality and reliability of the resulting knowledge graph. There exists a powerful bidirectional relationship between structured generation and graph operations. Structured generation enables reliable knowledge graph operations by maintaining entity consistency, relationship preservation, and error reduction. Meanwhile, the graph system enhances structured generation by providing contextual richness, relational understanding, temporal consistency, and reasoning support. This creates a virtuous cycle where each component strengthens the other. Agentic graph systems orchestrate multiple specialized components that must communicate effectively. Structured generation provides the "disciplined communication framework" required for component-to-component interaction. This is a "neural-symbolic bridge" in the architecture, enabling reliable information exchange between diverse system elements that would otherwise struggle with unpredictable outputs from each stage. Continuez en commentaire :

How do you predictably enforce the output format from an open source local model? ?? Use Pydantic and lmstudio-python ?? Learn more: https://lnkd.in/e6iCRZ9i