About Ax
We’re building the ultimate typescript framework for building agents and other complex LLM-powered workflows.
Motivation and vision
In the beginning, there was the LLM, a raw and powerful technology with endless potential but was hard to work with within your own software. Our first goal was to abstract out the LLM and have sensible defaults to build with any LLM while still using features like multi-modal, streaming, function calling, etc.
Next, we wanted to abstract out the prompting layer; we wanted it to be as fluid and flexible as writing code. We wanted automatic type enforcement, extensibility, composability, reuse, and more. We wanted prompting to be like coding. Our search brought us to the Stanford DSPy series of papers by Omar Khattab et al. Core to the paper was how prompts can be programs tunable by LLMs to become more efficient. This was fascinating, and we adopted several of these breakthroughs in our framework, such as prompt signatures, prompt tuning, assertions, etc.
Finally, we were ready to start building all the workflows we wanted on this solid foundation. Our first focus was RAG and the Agents both of which fully leverage the framework features like DSPy, streaming, mulit-modal, llm independence, etc.
We are excited to build more valuable layers on top of what we already built, Agent collaboration and Human-agent collaboration. We strongly believe that our current generation of LLMs is extremely powerful, and we can build fantastic things given the right framework and abstractions. Agentic workflows are the future of building with LLMs, and Ax is the best framework for that goal.
“I expect that the set of tasks AI can do will expand dramatically this year because of agentic workflows…”
Andrew Ng, Cofounder and head of Google Brain, former Chief Scientist at Baidu
Reach out
Features
- Full support for all top LLMs
- Multi-modal, function calling, streaming
- DSPy styled typed prompt signatures
- Build agents that can use other agents
- Automatic RAG, smart chunking, embedding, querying
- Convert docs of any format to text
- Output fields parsed and validated while streaming
- Multi-modal DSPy supported, use images with text fields
- Automatic prompt tuning using optimizers
- OpenTelemetry tracing / observability
- Production ready Typescript code
- Vertically integrated all parts work well together
- Lite weight, zero-dependencies