Local RAG

Retrieval-Augmented Generation, Done Right

RAG is how AI understands your codebase. But most implementations are cloud-based, slow, and generic. FrankenCoder's RAG is different:

Hybrid Search

Neither vector search nor keyword search is perfect alone:

• Vector search finds semantically similar code ("authentication logic" finds login functions)
• BM25 keyword search finds exact matches ("getUserById" finds that exact function)
• Hybrid combines both for the best results

FrankenCoder uses BGE-small for embeddings and BM25 for keywords, merged with intelligent ranking.

What Gets Indexed

• Code files: All programming languages
• Documents: Markdown, text, PDFs, Office files
• Conversations: Your chat history with AI
• Code symbols: Functions, classes, variables
• Dependency graphs: How files relate to each other

Per-Workspace Isolation

Each workspace has its own isolated index. This means:

• Project A's code doesn't leak into Project B's context
• Faster searches (smaller, focused indexes)
• Clean separation for client work
• Delete a workspace, delete its index

Real-Time Indexing

The indexing server runs continuously in the background:

• New files indexed automatically
• Changes detected and re-indexed
• Deleted files removed from index
• No manual "rebuild index" needed