MiniWiki - Repository Intelligence Interview

A minimal Next.js application for chatting with this specific repository.

Quick Start

1. Install Dependencies

npm install

2. Set Up Environment

Copy the environment template and add your Anthropic API key:

# Create .env file
echo "ANTHROPIC_API_KEY=your_api_key_here" > .env

Get your API key from Anthropic Console.

3. Run the Development Server

npm run dev

Open http://localhost:3000 to see the chat interface.

The Goal

Build a chatbot that can answer questions about this codebase.

The Challenge

Right now the chatbot has no information about the codebase and thus cannot answer any questions about it.

You could just concatenate the whole codebase into a string and pass that into the llm with the user message. That will work with this codebase but we don't want to do that.

Given that we don't want to stick the whole codebase into 1 LLM call, ou could give it tools like bash to read files and explore the code but then every time you ask a new question it'll have to do a bunch of exploratory work to look things up.

Instead, we want to build a simple RAG system instead that can retrieve the relevant files for any given user message from the codebase in 1 distinct step and pass those files to the final LLM call with the user message so that it has the context it needs to answer the question.

Question 1

What's the obvious limitation of passing in the whole codebase into the prompt?

Constraints

• Do not use a vector store
• You can assume 1 single file will never break the context window
• You should not use any external dependencies

Hint: You already have everything you need. There's a script at scripts/index-repo.ts you can run with npm run index. Start simple and iterate.

What's Already Built

✅ Chat Interface - A fully functional ChatGPT-style UI
✅ Streaming API - Backend with Anthropic Claude integration
✅ Index Storage - System to save and load your index
✅ CLI Runner - Infrastructure to run the indexing script

What You Need to Implement

1. Indexing Logic

• Analyze this codebase and create a searchable index
• Save it as JSON so it persists between chat sessions

2. Retrieval Logic

• Load the precomputed index
• Filter it based on the user's question
• Pass relevant code snippets to the LLM

Tips

• Start with the simplest approach that could work
• The precomputation step can use LLMs if helpful
• Test with real questions like "How does the chat API work?" or "What UI components are available?"
• Think about what information an LLM needs to answer code questions accurately