NodeSage

RAG-powered repo Q&A and code fixer CLI. Train it on any codebase, ask anything about the code, and fix issues — all locally using Ollama. Zero API costs, your code never leaves your machine.

Supports 40+ languages including JavaScript, TypeScript, Python, Go, Rust, Java, C/C++, Ruby, PHP, Haskell, Dart, Lua, SQL, Terraform, and many more.

Your Repo ──> Smart Chunker ──> Code Chunks ──> Embeddings ──> Vector DB
                                                                   │
                                                                   v
  You ──> "nodesage chat" ──> RAG Retriever ──> Ollama LLM ──> Answers / Fixes

1. Train — Scans your repo, chunks code by function/class boundaries, and embeds everything into a local vector DB at ~/.nodesage/
2. Chat — Ask questions in an interactive REPL. NodeSage retrieves relevant code via RAG and answers using a local LLM with streaming output
3. Fix — Auto-generates fixes with diffs, asks before applying, backs up originals

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

ollama serve                       # Keep this running
ollama pull qwen2.5-coder:7b      # Code LLM for Q&A and fixes (~4.7GB, 128K context)
ollama pull nomic-embed-text       # Embedding model for RAG (~274MB)

git clone https://github.com/vijaygupta18/NodeSage.git && cd NodeSage
npm ci
npm run build
npm link    # Makes 'nodesage' available globally

cd ~/code/my-project
nodesage train      # index the repo into ~/.nodesage/ (incremental on re-runs)
nodesage chat       # ask questions; /fix <file> to patch something from the REPL

Want to see the fixer in action without touching your own code? The repo ships with test-bad-code.js, a deliberately insecure Express app (SQL/command injection, hardcoded secrets, path traversal, sync I/O, unbounded cache, N+1 queries). Run nodesage fix ./test-bad-code.js and answer no at the apply prompt to keep the fixture as-is.

# Train on current directory
nodesage train

# Train on a specific repo
nodesage train ./path/to/repo

# Force full re-index (ignore cache)
nodesage train --force

Indexes all supported source files, embeds code chunks into a local vector DB at ~/.nodesage/. Supports incremental updates — only re-embeds changed files on subsequent runs.

nodesage chat

Starts an interactive session with streaming markdown output:

  NodeSage v2.0.0
  Ask anything about your trained codebase.
  Type /help for commands, /quit to exit.

  Model: qwen2.5-coder:7b

  > what does the chunker module do?
  ⠹ Thinking...

  The chunker module handles splitting source files into
  semantic chunks for embedding...

  > find security issues in the auth handler
  > how does the database connection pool work?

Chat commands:

• /fix <file> — Fix a file based on conversation context
• /model [name] — Switch LLM model mid-session (e.g. /model llama3.1:8b)
• /clear — Clear conversation history
• /context — Show what code chunks were retrieved for the last query
• /help — Show available commands
• /quit — Exit chat

nodesage fix ./src/app.js

Reviews the file, generates fixes, shows a colored diff, and asks before applying. Originals are backed up as .bak.

All settings are configurable via nodesage config, environment variables, or ~/.nodesage/config.json.

nodesage config

  NodeSage Configuration
  ~/.nodesage/config.json

  chatModel      qwen2.5-coder:7b (default)
  embedModel     nomic-embed-text (default)
  temperature    0.3 (default)
  maxTokens      4096 (default)
  topK           10 (default)
  chunkSize      40 (default)
  chunkOverlap   5 (default)
  contextWindow  20 (default)

# Switch to a different chat model
nodesage config chatModel llama3.1:8b

# Use a different embedding model
nodesage config embedModel mxbai-embed-large

# Adjust creativity (0.0 = deterministic, 1.0 = creative)
nodesage config temperature 0.5

# Increase max response length
nodesage config maxTokens 8192

# Retrieve more/fewer code chunks per query
nodesage config topK 15

# Reset a single key to default
nodesage config --reset chatModel

# Reset all config to defaults
nodesage config --reset

Override any setting without modifying the config file:

NODESAGE_CHAT_MODEL=llama3.1:8b nodesage chat
NODESAGE_EMBED_MODEL=mxbai-embed-large nodesage train
NODESAGE_TEMPERATURE=0.1 nodesage fix ./src/app.js
NODESAGE_TOP_K=20 nodesage chat

The --model flag on any command overrides the configured chat model for that session:

nodesage chat --model deepseek-coder:6.7b
nodesage fix ./src/app.js --model codellama:13b

NodeSage supports 40+ languages with smart boundary detection for accurate code chunking.

nodesage/
├── src/
│   ├── index.ts          # CLI entry point (train, chat, fix, config commands)
│   ├── config.ts          # Configuration management (file, env, defaults)
│   ├── types.ts          # Shared interfaces and types
│   ├── languages.ts      # Language detection + boundary patterns (40+ languages)
│   ├── chunker.ts        # Smart multi-language code chunker
│   ├── store.ts          # Vector store wrapper (Vectra, ~/.nodesage/)
│   ├── embedder.ts       # Ollama embedding (concurrent batch)
│   ├── retriever.ts      # RAG retrieval with file-aware context grouping
│   ├── llm.ts            # Ollama LLM wrapper (chat + streaming)
│   ├── trainer.ts        # Train command logic (full + incremental)
│   ├── chat.ts           # Interactive chat REPL with streaming markdown
│   ├── fixer.ts          # Fix command logic (standalone + interactive)
│   └── reporter.ts       # Terminal output (spinner, markdown rendering, diffs)
├── test-bad-code.js      # Intentionally bad Express app — fixture for trying `nodesage fix`
├── package.json
└── tsconfig.json

• Runtime: Node.js 20+ / TypeScript (glob@11 requires Node 20 or newer)
• CLI: Commander
• LLM: Ollama (any model — configurable)
• Embeddings: Ollama (nomic-embed-text default — configurable)
• Vector Store: Vectra (local, file-based)
• Output: Chalk (colored terminal with streaming markdown)

Issues and pull requests are welcome. See CONTRIBUTING.md for the dev setup, how to try changes against the bundled fixture, and what CI checks.

MIT