rag-rs

[!NOTE]

This software is released in alpha, so you can and should expect bugs. Please take a look at the Roadmap to see the plan to reach the first stable version

A Rust-native implementation of the RAG stack, using:

• PdfExtract for PDF extraction
• memchunk for chunking
• BM25 for embedding
• Qdrant for storing
• async-openai for LLM generation

Moreover, it can be served as an API server, usin:

• Axum as a web server framework
• tower-governor for rate limiting
• tower-http for CORS
• tracing and tracing_subscriber for logging

Installation and Usage

Install with cargo:

cargo install rag-rs

Install with npm:

npm install @cle-does-things/rag-rs@latest

load command

Parse, chunk and embed the documents in a given directory, and upload them to a vector store.

Usage

rag-rs load [OPTIONS] --directory <DIRECTORY> --qdrant-url <QDRANT_URL> --collection-name <COLLECTION_NAME>

Options

• -d, --directory <DIRECTORY>
  The path to the directory containing the files for the RAG pipeline. (required)
• --qdrant-url <QDRANT_URL>
  URL for a Qdrant vector store instance. If your Qdrant instance needs an API key, make sure that it is available as QDRANT_API_KEY in your environment. (required)
• --collection-name <COLLECTION_NAME>
  Name of the collection for the Qdrant vector store. (required)
• --chunk-size <CHUNK_SIZE>
  Chunking size. Default: 1024
• --cache-dir <CACHE_DIR> Directory where to cache the parsed file. Default: .rag-rs-cache/
• --cache-chunk-size <CACHE_CHUNK_SIZE> Chunk size for cached writes. Default: 1024 bytes
• --no-cache Deactivate read/write from cache. Default: active
• -h, --help
  Print help information.

Example

rag-rs --directory data/ \
    --chunk-size 2048 \
    --qdrant-url http://localhost:6334 \
    --collection-name test-data \
    --cache-dir cache/ \
    --cache-chunk-size 1048576

serve command

Serve the RAG application as an API server.

Usage

rag-rs serve [OPTIONS] --qdrant-url <QDRANT_URL> --collection-name <COLLECTION_NAME>

Options

• --qdrant-url <QDRANT_URL>
  URL of your Qdrant instance. If your Qdrant instance needs an API key, make sure that it is available as QDRANT_API_KEY in your environment. (required)
• --collection-name <COLLECTION_NAME>
  Name of the collection for the Qdrant vector store. (required)
• --openai-api-key <OPENAI_API_KEY>
  OpenAI API key. It is not advised to pass the key as an option to the CLI command: you should set it as the OPENAI_API_KEY environment variable.
• -p, --port <PORT>
  Port for the server to run on. Default: 8000
• --host <HOST>
  Host for the server to run on. Default: 0.0.0.0
• --rate-limit-per-minute <RATE_LIMIT_PER_MINUTE>
  Request rate limit per minute. Default: 100
• --cors <CORS>
  Allowed CORS origin (e.g. https://mydomain.com). Default: * (all origins allowed). While this argument has no effect for local development, it is advisable to set it for production deployments.
• --log-level <LOG_LEVEL>
  Logging level. Default: info
  Available values: info, debug, error, warning, trace
• --log-json
  Whether or not to activate JSON logging. Default: false (uses compact logging by default)
• -h, --help
  Print help information.

Example

rag-rs --qdrant-url http://localhost:6334 \
    --collection-name test-data \
    --host 127.0.0.1 \
    --port 3000 \
    --rate-limit-per-minute 30 \
    --cors "http://mydomain.com" \
    --log-leve info \
    --log-json

Limitations

• Currently supports only .pdf, .txt and .md files
• Does not go through the data directory recursively
• PDF extraction accounts only for text

Roadmap

To reach the first stable version, this software will first:

• Add a caching layer (v0.2.0-alpha)
• Introduce thorough testing (v0.2.1)
• Add an NPM-installable version (v0.2.1)

Moreover, for future releases, there will be:

• A programmatic API along with the CLI app, possibly both in Rust and Python
• Support for more text-based file formats, and possibly for more unstructured file formats