vai-workflow-code-search

Index, search, explain, and find duplicate code across any codebase using Voyage AI's code-optimized embedding models and MongoDB Atlas Vector Search. Transform any project into a semantically searchable knowledge base.

Install

vai workflow install vai-workflow-code-search

How It Works

1. Index — Scan, chunk, and embed all files in a local directory or remote GitHub repo into MongoDB Atlas Vector Search. Supports incremental refresh so you only re-index what changed.
2. Search — Search indexed code using natural language. Find implementations by intent, not just keywords.
3. Explain — Ask questions about the codebase and get AI-generated explanations grounded in the actual source code via RAG.
4. Find Similar — Paste a snippet or reference a file to detect duplicate or semantically similar code across the indexed codebase.

Pipelines

| Pipeline | Steps | Use Case | |----------|-------|----------| | Index & Search | index → search | Index a codebase then run a semantic search query | | Index & Explain | index → explain | Index a codebase then ask questions about it via RAG | | Full Code Audit | index → search → find-similar → explain | Index, search for patterns, find duplicates, and generate explanations |

Example Usage

# Index a local project and search
vai workflow run vai-workflow-code-search \
  --input source="/path/to/project" \
  --input query="where do we handle authentication timeouts?"

# Index a GitHub repo
vai workflow run vai-workflow-code-search \
  --input source="https://github.com/org/repo" \
  --input collection="my_codebase"

# Search indexed code (after indexing)
vai workflow run vai-workflow-code-search \
  --input query="How does the authentication middleware work?" \
  --input collection="my_codebase"

# Find similar code
vai workflow run vai-workflow-code-search \
  --input code="async function fetchUser(id) { return await db.users.findOne({ _id: id }); }" \
  --input collection="my_codebase"

Models Used

• voyage-code-3 — Code-optimized embeddings (default for source files)
• voyage-4-large — General-purpose embeddings (used for docs/markdown)
• rerank-2.5 — Reranker for improved search relevance

What This Teaches

• voyage-code-3 understands code semantics — variable names, control flow, API patterns
• Code snippets work as search queries, finding structurally similar code
• Incremental refresh detects changed files via content hashing and only re-embeds what changed
• RAG-powered explain uses retrieved code chunks as context for grounded, cited answers

License

MIT © 2026 Michael Lynn