Repopilot

Inspiration

AI coding tools have a “context blind spot.” They generate code without understanding the codebase—causing hallucinations, broken patterns, and unsafe changes.

We focused on solving this: not faster generation, but grounded, repo-aware guidance. RepoPilot acts as a Project Navigator, not a guesser.

What it does

RepoPilot AI is a “Chat with your Repo” assistant for onboarding and development.

• Input a GitHub repo
  
• Builds a semantic map of the codebase
  
• Answers high-level questions about structure and implementation
  
• Generates code aligned with existing patterns using real file references
  
• Refuses to answer when evidence is missing
  

How we built it

We used a modular RAG architecture:

• Frontend: Svelte
  
• Indexer: Python + GitHub API
  
• Pipeline:
  • Retriever → fetches relevant code
    
  • Planner → structures tasks
    
  • Verifier → ensures consistency with repo
    
• Grounding layer: enforces file citations and blocks unsupported outputs
  

Challenges we ran into

• Scaling context for large repos within token limits
  
• Reducing hallucinations while maintaining useful responses
  
• Teaching the model to reliably refuse instead of guessing
  

Accomplishments that we're proud of

• Strict grounding with real file citations (e.g., src/auth/utils.py)
  
• Verifier loop that prevents invalid imports and unsafe code
  
• Consistent pattern-matching with existing codebases
  

What we learned

Context matters more than model size. A grounded system outperforms a stronger model guessing blindly.

Developers value consistency with existing patterns as much as correctness.

What's next for RepoPilot

• GitHub bot for PR reviews and architectural checks
  
• Deeper dependency graphing across files and services
  

Built With

• github-api
• prompt-engineering
• python
• rag
• semantic-search
• static-analysis
• svelte
• vector-embeddings