RT by @marc_r_fischer: ✨Do you want to leverage LMQL generation/scoring abilities without having to write any LMQL code yourself? The new high-level Generations API introduced in LMQL 0.7 now enables you to do just that:

R to @lmqllang: The Generations API is available for all of LMQL's backends, including OpenAI, Transformers, AutoGPTQ, Azure and llama.cpp. This unifies LLMs access across your application, and interoperates with standard LMQL code seamlessly. Check the Docs: https://lmql.ai/docs/lib/generations.html

RT by @lmqllang: Our latest research shows an emerging privacy threat from LLMs - they can infer personal information from seemingly benign texts at scale! Check out our @WIRED feature: https://www.wired.com/story/ai-chatbots-can-guess-your-personal-information/ 👉Test your own inference skills against GPT-4 and learn more: https://llm-privacy.org/

@lmqllang could be a good candidate to control the conversation flow with #LLM and the output format. 😀 #buildinginpublic #indiehackers

R to @lmqllang: Check out the full Decorators documentation here: https://lmql.ai/docs/latest/language/decorators.html Also check out the Chat API, which leverages decorators to hide internal reasoning in chat applications: https://lmql.ai/docs/latest/lib/chat.html

R to @lmqllang: Lastly, a pre-processing decorator can be used to cache outputs per variable or use a different form of text generator for a specific intermediate output:

R to @lmqllang: Streaming decorators allow you to directly capture LLM output as it comes in, e.g. to stream output over the network to a client or website:

R to @lmqllang: Post-processing decorators allow you to transform generated data directly in-line during query execution. For instance, here, we transform a generated comma-separated list of items into a Python list object:

✨The newly introduced Decorators API in LMQL 0.7 allows you to run custom code during LLM inference. Decorator functions allow you to stream, pre and post process generated text during query execution. Read more 👇 or in the docs: https://lmql.ai/docs/language/decorators.html

Come and meet us at @ETH_AI_Center @ETH AI+X today. @lbeurerkellner will be giving an overview of how LMQL helps improve LLM safety, as part of the LLM safety workshop starting at 13:15. Also check out @mbalunovic segment on all the other exciting LLM safety research done @the_sri_lab. #LaunchAIXSummit2023

LMQL 0.7 comes with a built-in one-command way to deploy your LMQL query as an interactive chatbot. Make sure to check it out here: https://lmql.ai/docs/lib/chat.html

R to @lmqllang: And that's just the highlights! Check out the full release notes here https://lmql.ai/blog/posts/release-0.7.html, and make sure to have a look at the fully revised documentation at https://lmql.ai/docs/.

R to @lmqllang: 0.7 adds several new 💻backends: - OpenAI's gpt-3.5-turbo-instruct - http://replicate.com hosted models - the sentencepiece tokenizer, especially useful for llama.cpp models.

R to @lmqllang: Next to many new main-line features, 0.7 also includes several preview features: * LMQL Actions, a cross-model function calling layer * Regex Constraints * JSON/Dataclass Constraints

R to @lmqllang: The Chat API allows you deploy your query programs as fully functional Chatbots, via Web UI, WebSockets or HTTP SSE. The documentation also explains how to build LMQL chat applications, including chapters on serving, prompt injection defence and internal reasoning. Docs: https://lmql.ai/docs/lib/chat.html

R to @lmqllang: The new Generations API, provides a high-level API for using LMQL's core functionality without having to write any LMQL code yourself.

R to @lmqllang: just keeps getting better and better!