ZeroRAG is a modular, command-line Retrieval-Augmented Generation (RAG) pipeline designed for speed and simplicity. Instead of writing custom boilerplate for every new dataset, ZeroRAG allows you to transform a local folder of complex documents (PDFs, Word files) into a fully embedded vector database with a single command. Once your data is ingested, you can instantly query the database from your terminal to retrieve highly relevant text chunks—providing the perfect context window for LLM generation.

🚀 Install ZeroRAG

ZeroRAG is modular by design. The core package includes basic text processing (.txt) and a lightweight, file-based vector store. To keep your environment clean, you can opt-in to exactly the extra features your pipeline requires.

pip install zerorag

Optional Dependencies

📄 Document Parsers

Extend the core functionality to support additional file formats:

🧠 Embeddings

Use a different embedding provider:

Note: OpenAI embeddings require an API key. Set the OPENAI_API_KEY environment variable before use. See how to create an API key.

🗄️ Vector Stores

Use a different vector store backend:

🤖 LLMs

Use an LLM for answer generation with the ask command:

Note: OpenAI LLMs require an API key. Set the OPENAI_API_KEY environment variable before use. See how to create an API key.

⌨️ CLI Commands

ZeroRAG provides the following commands:

Ingest Documents

This command reads all supported documents from a source directory, splits them into manageable chunks, generates embeddings, and saves the resulting vector database.

Syntax:

zerorag ingest <SOURCE> [OPTIONS]

Arguments:

• SOURCE (Required): The folder containing your documents.

Options:

• --types (Default: txt): Comma-separated list of file types to parse (e.g., txt,pdf,docx).
• --chunk-size (Default: 1200): Maximum number of characters per chunk.
• --chunk-overlap (Default: 300): Number of overlapping characters between consecutive chunks.
• --embeddings (Default: fastembed): Embedding provider to use (fastembed, openai-small, or openai-large). The default uses FastEmbed (BAAI/bge-small-en-v1.5) locally with no API key required.
• --store (Default: inmemory): Vector store backend (inmemory or chromadb).
• --store-dir (Default: zerorag_store): Directory to persist the vector store.

Note: Parsing non-txt files requires installing the matching optional dependencies.

Example:

$ zerorag ingest ./docs --types txt,pdf --embeddings openai-small --store-dir ./my_store

🚀 Initializing ZeroRAG Ingestion...
   Source    : /home/user/docs
   Formats   : [txt, pdf]
   Embeddings: openai-small
   Store     : inmemory
   Store Dir : /home/user/my_store

📥 Loading documents...
✂️ Chunking documents...
🔢 Initializing embeddings...
💾 Storing vectors...
✅ Ingestion complete!

Query Documents

This command runs a similarity search against a vector store and returns the most relevant document chunks.

Syntax:

zerorag query <QUERY> [OPTIONS]

Arguments:

• QUERY (Required): The search query.

Options:

• --embeddings (Default: fastembed): Embedding provider to use (fastembed, openai-small, or openai-large). The default uses FastEmbed (BAAI/bge-small-en-v1.5) locally with no API key required. Must match the provider used during ingestion.
• --store (Default: inmemory): Vector store backend to load (inmemory or chromadb).
• --store-dir (Default: zerorag_store): Directory where the vector store was persisted during ingestion.
• --k (Default: 5): Number of top matching chunks to return.
• --full: Print the full content of each chunk instead of a truncated preview.

Example:

$ zerorag query "What are the main benefits of RAG?" --store-dir ./my_store --k 2

🔍 Querying vector store...
   Store     : inmemory
   Store Dir : /home/user/my_store
   Top K     : 2

📄 Result 1:
   Source: intro.pdf | Page: 2
   RAG combines retrieval with generation to ground LLM responses in
   real data, reducing hallucinations and improving factual accuracy...

📄 Result 2:
   Source: overview.txt
   The primary benefits include reduced hallucination, up-to-date answers
   from private data, and full traceability back to source documents...

✅ Query complete!

Ask a Question

This command retrieves relevant chunks from a vector store and sends them alongside your question to an LLM, returning a grounded answer based on your documents.

Syntax:

zerorag ask <QUESTION> [OPTIONS]

Arguments:

• QUESTION (Required): The question to ask.

Options:

• --llm (Default: openai-mini): LLM to use for generation (openai-mini or openai).
• --embeddings (Default: fastembed): Embedding provider to use (fastembed, openai-small, or openai-large). The default uses FastEmbed (BAAI/bge-small-en-v1.5) locally with no API key required. Must match the provider used during ingestion.
• --store (Default: inmemory): Vector store backend to load (inmemory or chromadb).
• --store-dir (Default: zerorag_store): Directory where the vector store was persisted during ingestion.
• --k (Default: 5): Number of top matching chunks to retrieve as context.

Note: The ask command requires installing the OpenAI optional dependency.

Example:

$ zerorag ask "What are the main benefits of RAG?" --store-dir ./my_store

🤖 Asking LLM...
   Question  : What are the main benefits of RAG?
   LLM       : openai-mini
   Embeddings: fastembed
   Store     : inmemory
   Store Dir : /home/user/my_store
   Top K     : 5

🔍 Retrieving relevant chunks...
💬 Generating answer...

📝 Answer:
   Based on your documents, the main benefits of RAG are:
   1. Reduced hallucination by grounding responses in real data
   2. Up-to-date answers from private data sources
   3. Full traceability back to source documents

✅ Done!

🐛 Reporting Bugs & Feature Requests

We are constantly looking to improve ZeroRAG. If you encounter a bug or have an idea for a new feature (like a new vector store or document loader), please let us know!

To report a bug or submit a feature request, visit the Issues page.