RAG-Enhanced Reviews

Boost review quality by automatically retrieving similar code from your indexed codebase and injecting it as additional context.

What is RAG?

Retrieval-Augmented Generation (RAG) is a technique that gives the AI access to your specific codebase's patterns, utility functions, and architectural conventions without needing to fit the entire repository into the model's context window.

When RAG is enabled, the extension finds existing code in your workspace that is semantically similar to the changes you're currently reviewing and provides it to the AI as a reference.

How It Works

The RAG implementation in Ollama Code Review consists of three main phases:

1. Indexing (The Knowledge Base)

The extension scans your workspace and builds a local searchable index:

• File Discovery: Finds files matching your includeGlob patterns while respecting excludeGlob (e.g., ignoring node_modules).
• Chunking: Splits source files into overlapping segments (default: 1,500 characters with 150-character overlap) to preserve context.
• Embeddings: Converts each code chunk into a high-dimensional vector using an embedding model (e.g., nomic-embed-text via Ollama).
• Storage: Persists the vectors and metadata in a flat JSON store (rag-index.json) within the VS Code global storage directory. This approach ensures compatibility with VS Code's sandbox and avoids complex native database dependencies.

2. Retrieval (The Search)

When you trigger a review:

1. Query Generation: The current git diff is converted into an embedding vector.
2. Semantic Search: The extension performs a Cosine Similarity search between the diff vector and all indexed code chunks.
3. Filtering: Results below the similarityThreshold (default: 0.65) are discarded. Snippets from files already included in the diff are also filtered out to avoid redundancy.
4. Ranking: The top-K most relevant snippets are selected for the final prompt.

3. Augmentation (The Review)

The retrieved snippets are injected into a dedicated section of the AI prompt titled "Related Code from Codebase". This allows the AI to:

• Identify if you are duplicating existing utility functions.
• Suggest usage of team-standard patterns found elsewhere in the repo.
• Catch inconsistencies with established architectural styles.

Setup & Configuration

Recommended Embedding Model

For the best results, we recommend using nomic-embed-text with local Ollama:

ollama pull nomic-embed-text

Fallback Mode

If no embedding model is available, the extension automatically falls back to a lightweight TF-IDF approximation. While less "semantic" than deep learning embeddings, it is highly effective at finding code with similar keyword and identifier usage.

Settings

Setting	Description	Default
rag.enabled	Enable RAG context injection	false
rag.indexOnStartup	Re-index workspace in the background on launch	false
rag.embeddingModel	Ollama model to use for vectors	nomic-embed-text
rag.maxResults	Number of snippets to include in reviews	5
rag.similarityThreshold	Precision filter (0.0 to 1.0)	0.65

Commands

• Ollama Code Review: Index Codebase: Manually trigger a full re-index of the current workspace.
• Ollama Code Review: Clear RAG Index: Delete the local vector store to save space or force a clean rebuild.