What is Lune AI?
And why are we building it?
There are many AI tools for code popping up in recent times, ranging from AI-integrated VS Code forks to general foundational LLMs that have become decent at writing python scripts. There's no doubt that many of these tools boost productivity for engineers. However, we're not at the point where we can fully entrust AI to write all our code for us, and that time could be a long ways away. A problem to be solved now, with the advent of LLM commoditization, is getting accurate, up-to-date technical knowledge, and fast.
Before ChatGPT, Claude, etc, if you had a question about an API or library, you had two choices. One, actually read the documentation and research the problem. Two, ask another human being (hopefully after actually still reading the documentation). But now, you have a third choice. Ask ChatGPT and get an answer instantly. Even better, piece together a badly worded question and paste in your code, and get it to output a proposed solution that you can conveniently paste into your IDE.
Except... you have no idea if the answer was hallucinated or not.
"Write me a Python script to parse this JSON file"
General LLMs will handle this type of query very well. In fact, it can handle much more complicated queries fairly decently. But what about other types of questions?
"Summarize the newest changelog for me"
Nope - what changelog?
"How do I get started with streaming runnables with Langchain?"
Good luck with this one.
LLMs are static in nature. Any time your query requires specific, recent knowledge, chances are the LLM will hallucinate. Open AI is integrating web-search and Perplexity does pull context from websites, but you still don't have fine control over exactly which knowledges sources its pulling from. And the results are still far from perfect. To address these challenges with technical queries, we are building Lune.
So what are we building at Lune AI?
Users can train individual "Lunes" on a technical knowledge base. This knowledge base can be composed of multiple different sources, such as API references, documentation, READMEs, and more. Each Lune is a conversational AI, similar to ChatGPT, except it knows when and how to pull from the user-defined knowledge base.
The best part, a Lune can be trained in a few minutes. Create an account and explore public Lunes that other users have already created, and if you can't find one that you're looking for, then try making one yourself!
Eventually, we hope to make both individual public and private Lunes available through an API, so users and teams can integrate them into their own products.
Guess it's cool that anyone can create a Lune, but is it just an LLM + retrieval?
Here's where things get more exciting. In the next few weeks, we are gonna be launching community-oriented features. These include:
- Any user can contribute to the knowledge base of a public Lune (think Wikipedia for Lunes)
- Not satisfied with an answer from a Lune? Automatically create a community post about this conversation thread in seconds. If there is a verified community answer, Lune will automatically ingest that community interaction into it's knowledge base, to improve future outputs.
The vision: Convenience of LLMs + reliability of community backed knowledge.
Going into the future, we don't think that the first place you go to with a technical question should be a real person. We also don't think it should be a general foundational LLM, as many engineers are doing now. We think it should be a platform that provides the tool capable of answering 90% of queries accurately, fine-tuned over thousands of real interactions, while providing a robust network of human-reinforced knowledge for the remaining 10%. We are building that platform with Lune.
Questions? Comments? Disagree? Reach us at sam@trylune.ai and josh@trylune.ai
