Skip to main content

​
How querying works

Once your repository is indexed, you can ask questions in natural language. RepoRAGX uses a RAG (Retrieval-Augmented Generation) pipeline to answer:
1

Query embedding

Your question is converted into a vector embedding using the same Sentence Transformers model (all-MiniLM-L6-v2) used during indexing.
2

Similarity search

ChromaDB performs a cosine similarity search to find the top-K most relevant code chunks (default: 5 chunks).
3

Context retrieval

The most relevant code snippets are retrieved along with their file path metadata and similarity scores.
4

LLM generation

The retrieved context and your question are sent to the Groq LLM (llama-3.3-70b-versatile by default), which generates a context-aware answer.

​
Asking questions

At the main prompt, type your question and press Enter: 
Ask anything ('exit' to quit): Where is authentication implemented?
You’ll see diagnostic output showing the retrieval process: 
Running RAG for query: Where is authentication implemented?
Retrieving documents for query: 'Where is authentication implemented?'
Top K: 5, Score threshold: 0.0
Retrieved 5 documents (after filtering)
Then the LLM’s answer: 
> Authentication is implemented in server/controllers/authController.js. The file contains functions for user registration, login, and token verification using JWT tokens.

​
Types of effective queries

RepoRAGX works best with specific, code-focused questions:

​
Location queries

Ask where specific functionality is implemented: 
Where is error handling implemented?

​
Implementation queries

Ask how something works: 
How does the text splitting work?

​
Discovery queries

Find out what exists in the codebase: 
What API endpoints are available?

​
Explanation queries

Get explanations of existing code: 
What does the GitHubCodeBaseLoader class do?

​
Understanding retrieval results

​
Retrieval parameters

The RAG retriever uses these default parameters (defined in src/rag/rag_retriever.py:7):
  • top_k: 5 - Number of most relevant chunks to retrieve
  • score_threshold: 0.0 - Minimum similarity score (0-1 range)
Results are ranked by similarity score, where similarity_score = 1 - distance.

​
Context provided to LLM

Each retrieved chunk includes:
  • File path: The source file location (e.g., src/rag/github_codebase_loader.py)
  • Content: The actual code snippet
  • Similarity score: How relevant the chunk is to your query
  • Rank: Position in the results (1-5 by default)
The LLM receives all this context formatted as: 
--- File: src/rag/groq_llm.py ---
[code content]

--- File: src/rag/rag_retriever.py ---
[code content]

​
No results found

If your query doesn’t match any code in the repository, you’ll see: 
No documents found
> No relevant context found to answer the question.
Try rephrasing your question or using different keywords that might appear in the actual code.

​
Continuous interaction

After each answer, you’ll see the prompt again: 
Ask anything ('exit' to quit):
You can ask as many questions as you want in a single session. Each query is independent and uses the same indexed vector store.

​
Exiting the session

To end your session and exit the application, type exit: 
Ask anything ('exit' to quit): exit
The comparison is case-insensitive and strips whitespace, so EXIT, Exit, and exit all work.

​
Best practices

Use terms that are likely to appear in the actual code. If you’re looking for database functionality, use terms like “database”, “connection”, “query” rather than vague terms.
Break complex questions into smaller, focused queries. Instead of “How does the entire RAG pipeline work?”, ask about specific components like “How are documents chunked?” then “How are embeddings generated?”
If you know a file exists, mention it: “What does the main.py file do?” or “Explain the GitHubCodeBaseLoader class”
If an answer isn’t clear or complete, ask follow-up questions to drill deeper into specific aspects.

Build docs developers (and LLMs) love

Get started for freeTalk to us