Dastavej RAG

A simple Python framework for building multi-user Retrieval-Augmented Generation (RAG) applications.

dastavej-rag provides PDF ingestion, text chunking, embeddings, Qdrant vector storage, user-isolated semantic retrieval, duplicate-document detection, source tracking, and LLM-based question answering.

Features

• PDF ingestion with PyMuPDF
• Plain-text ingestion
• Automatic text chunking with overlap
• Sentence Transformer embeddings
• Qdrant vector storage
• Persistent local vector database
• User-isolated retrieval using user_id
• Duplicate PDF detection using SHA-256
• Semantic similarity search
• Gemini LLM integration
• Custom LLM support
• Source and page tracking
• Document deletion
• User-data deletion
• Structured RAGResponse
• Structured Source objects

Installation

pip install dastavej-rag

Gemini Setup

Set your Gemini API key:

export GEMINI_API_KEY="your-api-key"

Never commit API keys to source control.

Quick Start

import dastavej_rag as dr

rag = dr.RAG(
    llm=dr.GeminiLLM()
)

result = rag.add(
    "company_policy.pdf",
    user_id="user_123"
)

print(result)

response = rag.ask(
    "What is the leave policy?",
    user_id="user_123"
)

print(response.answer)

for source in response.sources:
    print(source.source)
    print(source.page)
    print(source.score)

Add Plain Text

rag.add(
    "Qdrant is a vector database.",
    user_id="user_123"
)

You can also use the explicit API:

rag.add_text(
    text="Qdrant is a vector database.",
    user_id="user_123"
)

Retrieve Chunks

results = rag.retrieve(
    query="What vector database is used?",
    user_id="user_123",
    top_k=3
)

for result in results:
    print(result["text"])
    print(result["score"])

Delete a Document

rag.delete_document(
    document_id="document-id",
    user_id="user_123"
)

Delete User Data

rag.delete_user_data(
    user_id="user_123"
)

Custom LLM Providers

You can implement your own provider by extending BaseLLM:

import dastavej_rag as dr


class MyLLM(dr.BaseLLM):

    def generate(self, prompt: str) -> str:
        # Call your preferred LLM here.
        return "Generated answer"


rag = dr.RAG(
    llm=MyLLM()
)

This allows Dastavej RAG to work with other LLM providers.

RAG Pipeline

PDF / Text
    ↓
Text Extraction
    ↓
Chunking
    ↓
Embeddings
    ↓
Qdrant
    ↓
user_id Filtering
    ↓
Semantic Retrieval
    ↓
Context Construction
    ↓
LLM
    ↓
RAGResponse
    ├── answer
    └── sources

Multi-User Applications

Every stored chunk contains a user_id.

Retrieval and deletion operations filter using that user_id, allowing applications to logically isolate documents belonging to different users.

Authentication and authorization remain the responsibility of the application using dastavej-rag. Applications should provide a trusted user_id derived from their authenticated user/session rather than trusting arbitrary client input.

Duplicate Detection

PDF files are hashed using SHA-256.

Uploading the same PDF again for the same user returns a duplicate result instead of embedding and storing the document again.

The same document may still be independently stored for another user.

Local Storage

By default, Qdrant data is persisted locally under:

.dastavej_qdrant/

Add this directory to .gitignore.

Development

Install development dependencies:

pip install -e ".[dev]"

Run tests:

python -m pytest -v

Current Version

0.1.0

Dastavej RAG is currently an alpha release.

License

MIT