Kedar Chandrayan

Looking inside a @sirji_ai agent and recipe with @TechieKedar.

LLMs (and some new AI tools) do a decent job at generating code for a new project. But they struggle with existing codebases. Why? 1. Existing Architecture: New projects don't have the baggage of existing architecture and design patterns that AI needs to understand and fit into. 2. Project-Specific Conventions: Existing projects come with unique coding standards and conventions. LLMs fail to follow these without explicit and detailed guidance. 3. Contextual Awareness: Understanding the full context of the existing codebase, including why certain decisions were made, is critical. LLMs lack this context. But then if you really want to build a tool that developers will use for their everyday dev tasks, it needs to be able to work with existing codebases and not just create new shiny projects. It's hard but not impossible. That’s what we are trying to do with @Sirji_ai.

How is #Sirji different from github copilot workspace? So while I haven’t got access to github copilot workspace yet, the main difference is that with Sirji you can write custom agents. (1/n) 🧵

We have been using #Sirji to create new APIs and test cases for Raised (one of our side projects). This is an existing repository with 100+ files. It took @TechieKedar 10 minutes while working with Sirji to write a new API and the corresponding test cases. It took 20 minutes more to manually review the code and to make two minor edits. Without Sirji, this task would have taken him 5-6 hours. That’s a 10x productivity boost!

What’s going on with you, #Sirji? A question I’ve been asking our AI software development agent. After launching Sirji’s VS code extension (which really is a breeze to install), I’ve been collating all the feedback from early users - both good and bad. And here’s the compilation of all the issues that have frustrated our early users. * Sirji says it is done solving the problem even when it hasn’t fully implemented the expected functionalities. * It doesn’t yet have the capability to debug web errors and hence relies on users to report errors by copying and pasting them manually. * Sirji often says "Outside of my capability" for tasks that are very well within its capabilities (like installing MongoDB). * At times, Sirji installs system-wide packages like Python without the user’s explicit approval. * A serious issue is when Sirji occasionally attempts to create files and folders outside the designated workspace. Our next milestone is all about making Sirji reliable - making Sirji an agent you can trust with your problem.

It’s 1 am, but what has to be shipped can’t be held up. Here’s the first version of #Sirji as a VS Code extension. Install it in under 5 seconds & check it out for yourself.

#Sirji Update: Possible solutions to equip the Executor with debugging capabilities: 1. Use tee command 2. Use “vscode.window.onDidWriteTerminalData” 3. Use just the stdout stream Are there any other solutions that we can explore? For more details: github.com/sirji-ai/sirji…

Do you know why “Enter number: ” is not shown on my terminal (i.e. stdout), but instead in error.log (i.e. stderr)? File user_input.py: num = input(“Enter number: “) print(num) command:  python user_input.py 2> error.log #sirji #buildinpublic

#sirji #buildinpublic I wanted to share my proposal to enhance Sirji's capabilities to navigate and work with GitHub repositories. Strategy: Initial shortlisting of files, guided by directory structure, followed by incremental deep dive. Details 👉 github.com/sirji-ai/sirji…

Attn devs: Sirji's dogfood release is done. What's in dogfood: * Chat interface for the user to enter the problem statement. The user can interact with Sirji by providing change requests and enhancements. * Message protocol implemented for various agents - Planning Agent, Coding Agent, Research Agent, and Executor Agent. * The Planning Agent: Takes in a problem statement and breaks it down into steps. * The Coding Agent: Sequentially generates code to solve the problem programmatically. * The Research Agent: Utilizes RAG and gets trained on URLs and search terms. Later, it can use the training knowledge to answer the questions asked by the Coding Agent. * The Executor Agent: Responsible for creating the code files, running commands, and installing dependencies. You can see what's possible with the dogfood release in the demo videos. Link in my comment. Now onwards to the next release: Visual Studio Code IDE as the user interface.

Prompt for Sirji: Create an express app. Use PostgreSQL as database and Sequelize as ORM. Create migration files and model files by understanding the DBML file. Create API routes in modular files by understanding the OpenAPI specs file. See the magic happen!

#sirji #buildinpublic Sirji’s chat GUI got built today. It helps the user to enter the problem statement and give feedback on the solution.

And here we go… Sirji’s first successful run. 🚀 I wouldn’t say we are done with dogfood release yet. This was just one run. We will be testing with a variety of different problem statements. The main functionality remaining for this release is adding debugging capabilities (including user feedback & interaction).

#sirji We have standardized the system prompts for each of the agents, and created a simple protocol for inter-agent communication. Here's @sunilkhedar's update on the messages & prompts packages.

@TechieKedar Alongside, @rachinkapoor is already exploring Sirji’s eventual IDE and why he’s drawn towards solutions like codespace.

#sirji #buildinpublic @TechieKedar talks about what you can expect from the dogfood version (ETA: this Friday) of Sirji. It’s going to be a basic command line interface with terminal windows with logs from various agents.

#sirji #buildinpublic Here we go! Today was about releasing the first version of the Researcher module which enables Sirji to gain new knowledge and infer from it.

Here’s the PR that got Researcher to the main branch: github.com/sirji-ai/sirji… So hey, time to get your hands dirty. Run the Researcher on your machine by following these steps: github.com/sirji-ai/sirji…

Sirji is everywhere. A fun jab among friends. A chuckle on a zoom call. Much more than respect. It's our vibe. Inspired by Devin, Sirji is an open-source AI software agent. Being built in public starting today by True Sparrow.

Really cool work from @RLanceMartin on the @langchain around multi needle He found that: * Yes, more needles is harder * Yes, longer context is harder * Yes, increased reasoning is harder * The top of the document has less recall than the 2nd half