Probe

Project-scoped RAG for AI coding agents. Hybrid dense + BM25 retrieval, localhost-first.

Install

# 1. Start services (pick your hardware)
docker compose --profile lite up -d      # Laptops / modest GPUs
docker compose --profile balanced up -d  # Mid-range GPUs (~8GB VRAM)
docker compose --profile pro up -d       # High-end GPUs (~16GB+ VRAM)

# 2. Initialize your project
cd /path/to/your/project
probe init

# 3. Start the server
probe serve --watch

That's it. Add to your editor's MCP config and go.

Presets

Preset	Embedding Model	Reranker	Use Case
lite	Qwen3-Embedding-0.6B	—	Laptops, CPU, quick iteration
balanced	Qwen3-Embedding-4B	Qwen3-Reranker-0.6B	Good quality, reasonable VRAM
pro	Qwen3-Embedding-8B	zerank-2	Best quality, needs big GPU

Configuration

Environment Variables

Variable	Default	Description
QDRANT_URL	http://127.0.0.1:6333	Vector database
TEI_EMBED_URL	http://127.0.0.1:8080	Embedding service
RERANKER_URL	http://127.0.0.1:8083	Reranker (balanced/pro only)
TEI_IMAGE_TAG	1.8	TEI Docker image version

GPU Architecture

Override TEI_IMAGE_TAG for your GPU:

GPU	Tag
A100 (Ampere 80)	1.8 (default)
A10/A40 (Ampere 86)	86-1.8
RTX 4000 series (Ada)	89-1.8
T4/RTX 2000 (Turing)	turing-1.8
CPU	cpu-1.8

TEI_IMAGE_TAG=turing-1.8 docker compose --profile lite up -d

CLI

probe init              # Initialize workspace
probe serve --watch     # Run MCP server + file watcher
probe scan              # Force reindex
probe prune             # Clean orphan workspaces
probe doctor            # Health check

Editor Setup

Claude Code

Create .mcp.json in your project root:

{
  "mcpServers": {
    "probe": {
      "command": "probe",
      "args": ["serve", "."]
    }
  }
}

Cursor

Add to your MCP settings (Settings > MCP):

{
  "mcpServers": {
    "probe": {
      "command": "probe",
      "args": ["serve", "."]
    }
  }
}

Codex

Add to ~/.codex/config.toml or project .codex/config.toml:

[mcp_servers.probe]
command = "probe"
args = ["serve", "."]

Troubleshooting

OOM on startup: Use a lighter preset or reduce batch size.

TEI won't start on CPU: Use TEI v1.8.2+ (cpu-1.8). Versions 1.8.0-1.8.1 have MKL bugs with Qwen3.

V100 not working: V100 (compute 7.0) isn't supported. Use CPU mode or upgrade GPU.

Switching presets: Requires full reindex (each preset uses different vector dimensions).

Two-Computer Setup

GPU box + laptop? Use SSH tunnels instead of exposing services:

# On laptop: forward ports from GPU box
ssh -N \
  -L 6333:127.0.0.1:6333 \
  -L 8080:127.0.0.1:8080 \
  -L 8083:127.0.0.1:8083 \
  user@gpu-desktop

Now your laptop talks to 127.0.0.1 — no auth needed.

License

MIT