Build your second brain with Khoj AI — High Signal AI #2

Harness the Power of Khoj with Obsidian, monitor agents on AgentOps

The best of AI Engineering resources summarised!

[Original Post] here:

Build your second brain with Khoj AI - High Signal AI #2

This week’s action-prompting insights:

1. Top find of the week 🏆: Khoj
2. Developer tool of the week
3. GitHub repositories you must check out!
4. What people are building with AI!
5. AI startups that raised funds
6. AI tutorials and videos on YouTube worth your time
7. What I’m reading this week!
8. Weekly updates from tools, libraries and hubs

Top find of the week 🏆: Khoj

https://github.com/khoj-ai/khoj

Khoj is an open-source, personal AI application designed to serve as a second brain.

It answers questions from online sources or personal notes using foundation models or private local LLMs.

Users can self-host locally or use a cloud instance. Khoj integrates with platforms like Obsidian, Emacs, a desktop app, web, and WhatsApp, offering features such as semantic search, document sharing, and personalized AI agents.

The project encourages community contributions and supports self-hosting.

You can contribute to the project by working on good first issues like this one.

I tried it integrating Khoj on my Obsidian notes, worked like a charm:

Build a second brain with Khoj 🧠 #ai #obsidian #plugins #productivity #engineering #notes

Developer tool of the week: AgentOps

AgentOps is a platform designed to simplify and automate the deployment, management, and monitoring of AI agents.

It offers tools for deploying AI models, managing workflows, and ensuring smooth operations with features like real-time monitoring and alerts.

GitHub repositories you must check out!

RAG Search API — https://github.com/thinkany-ai/rag-search (677 stars) — This repository “rag-search” by thinkany.ai provides a RAG (Retrieval-Augmented Generation) Search API.

• It enables users to perform semantic search queries using a combination of retrieval and generation techniques.
• The API supports integration with various search providers, like Google, and employs models such as OpenAI’s GPT-3.5-turbo for query processing.
• The setup involves configuring environment variables, installing dependencies, and running a FastAPI server
• The API endpoints include basic ping and RAG search functionalities, allowing for detailed and filtered search results.

SWE-agent — https://github.com/princeton-nlp/SWE-agent (10.8k stars) — Software Engineering Agents are being loved by many. We checked out Devika last week, this week’s most loved and trending agent is SWE-agent.

• SWE-agent takes a GitHub issue and tries to automatically fix it, using GPT-4, or your LM of choice. It solves 12.29% of bugs in the SWE-bench evaluation set and takes just 1.5 minutes to run.

What people are building with AI!

Suno-API — (1.2k stars) — The “Suno-API” repository is an unofficial API based on Python and FastAPI for interacting with Suno AI services.

• It supports generating songs, lyrics, and more. Key features include automatic token maintenance, keep-alive functionality, and asynchronous operation.

Visualise protein in 3D on HuggingFace Space (claiming to be alternative of AlphaFold 3) — The “proteinviz” project is an open-source alternative to AlphaFold3 for protein structure prediction.

• It uses Facebook’s ESMFold model to predict 3D structures from amino acid sequences, generating PDB files for visualization.
• The tool leverages biopandas, plotly, graphein, and PIL for rendering, and provides an intuitive Gradio interface for user interaction. Designed for research purposes, it requires a good GPU to run efficiently.

A useful and short course you can complete this week:

Multi AI Agent Systems with crewAI — The “Multi AI Agent Systems with crewAI” course by DeepLearning.AI teaches designing and prompting teams of AI agents to automate complex tasks.

This course is taught by the founder of CrewAI himself.

Using the crewAI library and you can automate multi-step business processes like resume tailoring, customer support, and financial analysis.

The course covers role assignment, memory, tools, focus, error handling, and cooperation among agents. It is ideal for those with basic coding skills and some prompt engineering experience.

AI startups that raised funds

1. Scale AI: Scale AI raises $1B at $13.8B valuation in round led by Accel.
2. Suno AI: Suno raises $125M to enable anyone to make music with AI.
3. Angel AI: Angel AI has raised $4.75M in seed funding led by Cortical Ventures, with participation from Village Global and other investors.

AI tutorials and videos on YouTube worth your time

• LangGraph + CrewAI: Crash Course for Beginners — An introduction to using LangGraph and CrewAI for building and managing AI agents. It covers the basics of setting up the tools, writing code to deploy AI models, and integrating these models into workflows.
• RAG for Long Context LLMs discusses the implementation of Retrieval-Augmented Generation (RAG) in the context of long context language models (LLMs).
• It highlights how RAG can handle context windows exceeding 1 million tokens, enabling more comprehensive and efficient information retrieval.
• If you’re interested in learning about RAG from scratch, I suggest you check out their complete playlist : RAG From Scratch.

What I’m reading this week!

• Kyle Corbitt gave a talk on 10 commandments to deploy fine-tuned models in production — here is the complete slide deck that has a bunch of tenets that will guide you on when you should fine-tune and what are the best practices.
• Let’s talk about LLM evaluation — The Hugging Face blog post on LLM evaluation explores various methods for evaluating large language models (LLMs), including automated benchmarking, human judgments, and using models as judges.
• Building RAG on complex PDFs.

Weekly updates from tools, libraries and hubs

• Microsoft has released Phi-3 small (7B) and medium (14B) models under the MIT license, with instruct versions supporting up to 128k context.
• Additionally, the Phi-3-vision model, containing 4.2B parameters, has been introduced. ML Engineers have been saying that Phi-3-vision surpasses larger models like Claude-3 Haiku and Gemini 1.0 Pro V in visual reasoning tasks.

I capture all such AI powered developer friendly resources in my newsletter called High Signal AI. Subscribe and I’ll deliver these straight to your inbox.