RAG Stash

Perform RAG flexibly and easily with an LLM of your choice

About

Ragstash is a CLI utility to help retrieve RAG (Retrieval Augmented Generation) context from your documents which can then be fed into an LLM of your choice. It is a simple tool meant to be integrated into your RAG workflow. You could even ask a local LLM (that you trust) to fetch RAG context by using the tool itself while a server is running.

Installation

Available on PyPI

recommended:

pipx install ragstash or uv tool install ragstash

You should then have ragstash command available.

Usage

First, cd into your documents folder, then:

1. Use ragstash init to initialize (processes documents, generates embeddings, and saves vector db)
2. Use ragstash serve to start the server used for querying
3. (In a separate terminal) Use ragstash get <QUERY> to print out the RAG context. This output can be piped (|) into an LLM or saved in a file.

Example Usage

Simple:

> cd path/to/your/documents
> ragstash init
> ragstash serve

# In another terminal
> ragstash get "What do the leaked files say about the doings of Celebrity Celebface?" | claude

Advanced:

> cd path/to/your/documents
> ragstash init \
  --sentence-transformer "sentence-transformers/all-MiniLM-L6-v2" \
  --chunk-size 1000 \
  --name "LargeChunks"
> ragstash serve --name "LargeChunks"

# In another terminal
> ragstash get "What do the leaked files say about the doings of Celebrity Celebface?" \
  --retrieval-query "Things Celebrity Celebface has done" \
  --chunks 10 \
  --message "I am doing RAG, here are some chunks of info ^^" \
  | claude

Options

> init PTH
    --sentence-transformer  # model to use for embedding of chunks
    --chunk-size            # size of chunks (chars)
    --chunk-overlap         # overlap of chunks (chars)
    --redo                  # overwrite existing vault
    --name                  # name to give initialization (appended to folder name: ".rag_{NAME}")

> serve PTH
    --name                  # which initialization to use
    --unload-timeout        # time after which sentence transformer mode unloads (0 means never unloads)
    --port

> get YOUR_QUERY
    --chunks                # number of chunks to fetch
    --retrieval-query       # query to use when doing "dumb" retrieval of data (given to sentence transformer)
    --message               # message to give LLM
    --port                  # port to fetch
    --ip-addr               # default `localhost`
    --chunks-as-json        # instead of getting RAG formatted message, get retireved chunks as json
    --file                  # file where to save retrieved context

> update PTH
    --name                  # name of vault

Upcoming features

• update mode so that you can add or update documents and only scan the changes.