ROOT /

History

Enable CI/CD by adding .onedev-buildspec.yml

src/qdrant_indexer	Loading last commit info...
tests
.gitignore
.python-version
README.md
docker-compose.yml
flake.lock
flake.nix
pyproject.toml
uv.lock

README.md

Qdrant Indexer

CLI tool for indexing documentation and code into Qdrant for semantic search.

Qdrant Server Setup

docker compose up -d

This starts Qdrant with REST API on localhost:6333 and gRPC on localhost:6334.

Installation

Using uv (recommended)

uv pip install -e .

Using Nix

# CPU
nix develop --impure

# GPU/CUDA (Linux only)
nix develop .#cuda --impure

Usage

Index files

# Index markdown/text/pdf files
qdrant-indexer index ./docs -c my-docs

# Index code
qdrant-indexer index ./src -c my-code -p "**/*.py" -p "**/*.rs"

# With GPU acceleration
qdrant-indexer index ./docs -c my-docs --gpu

# Full re-index (default is incremental)
qdrant-indexer index ./docs -c my-docs --full

# Specify chunking strategy (see Chunking Strategies section)
qdrant-indexer index ./docs -c my-docs --chunker markdown

Other commands

# Check indexing status
qdrant-indexer status ./docs

# List collections
qdrant-indexer list-collections

# Delete collection
qdrant-indexer delete-collection my-docs

# List available embedding models
qdrant-indexer list-models

Chunking Strategies

The --chunker option controls how documents are split into chunks for indexing:

Strategy	Description	Best For
`auto`	Automatically selects optimal chunker based on file type	General use (default)
`recursive`	General-purpose chunker using paragraph/line/sentence splitting	Plain text, mixed content
`fixed`	Simple fixed-size chunks without semantic awareness	When consistency matters
`markdown`	Markdown-aware chunking that splits on header boundaries	Markdown documentation
`html`	HTML-aware chunking that splits on semantic tags (article, section, etc.)	Web pages, HTML docs
`semantic`	Embedding-based chunking that splits at semantic boundaries	Highest quality retrieval
`code`	Code-aware chunking that respects symbol boundaries (functions, classes)	Source code files

Usage Examples

# Auto-select chunker based on file type (default)
qdrant-indexer index ./docs -c my-docs --chunker auto

# Use markdown-aware chunking for documentation
qdrant-indexer index ./docs -c my-docs --chunker markdown

# Use semantic chunking for better quality (slower)
qdrant-indexer index ./pdfs -c papers --chunker semantic

# Force recursive chunking for all files
qdrant-indexer index ./docs -c my-docs --chunker recursive

# Use code chunking for source files
qdrant-indexer index ./src -c my-code --chunker code

Performance Notes

recursive, fixed: Fastest (baseline)
markdown, html, code: ~10% slower than recursive
semantic: 10-20x slower (uses embedding model for boundary detection)

Troubleshooting

Chunks too small/large: Adjust --chunk-size and --chunk-overlap parameters
Poor retrieval quality: Try semantic chunker or increase overlap
Slow indexing: Use recursive or fixed instead of semantic
Code symbols split incorrectly: Use code chunker for source files

MCP Setup

Claude

Add to your Claude MCP configuration (~/.claude.json or project .mcp.json):

{
  "mcpServers": {
    "qdrant-docs": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "mcp-server-qdrant",
        "--qdrant-url", "http://localhost:6333",
        "--embedding-model", "sentence-transformers/all-MiniLM-L6-v2"
      ]
    }
  }
}

GitHub Copilot

Add to .vscode/mcp.json:

{
  "servers": {
    "qdrant-docs": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "mcp-server-qdrant",
        "--qdrant-url", "http://localhost:6333",
        "--embedding-model", "sentence-transformers/all-MiniLM-L6-v2"
      ]
    }
  }
}

For both setups, ensure you use the same embedding model that was used during indexing.