rag-kit

rag-kit is a simple, modular Python library for building PDF-based RAG applications with conversational memory and flexible LLM provider support.

It is designed to hide most of the LangChain complexity behind a clean API:

from ragkit import PDFRAG

rag = PDFRAG("data/sample.pdf")
print(rag.ask("What is LangChain?"))

---

Features

• PDF-based RAG
• Conversational chat with session memory
• Follow-up handling for queries like:
  • hindi m batao
  • tell me in english
  • what did I ask earlier?
• Query rewriting for better retrieval
• Source return support
• Configurable chunking and retrieval
• Multiple LLM provider support:
  • Sarvam (default)
  • OpenAI
  • Anthropic / Claude
  • Custom LangChain-compatible chat models

---

Installation

Basic install

pip install rag-kit

Optional provider extras

pip install "rag-kit[openai]"
pip install "rag-kit[anthropic]"
pip install "rag-kit[all]"

Local development install

pip install -e .

---

Environment Variables

Create a .env file in your project root:

SARVAM_API_KEY=
OPENAI_API_KEY=
ANTHROPIC_API_KEY=

An example template is provided in .env.example.

---

Quick Start

Stateless Q&A

from ragkit import PDFRAG

rag = PDFRAG("data/sample.pdf")
answer = rag.ask("What is memory?")
print(answer)

Chat with memory

from ragkit import PDFRAG

rag = PDFRAG("data/sample.pdf")

session_id = "user1"

print(rag.chat("What is memory?", session_id=session_id))
print(rag.chat("hindi m batao", session_id=session_id))
print(rag.chat("tell me in english", session_id=session_id))

Return sources

from ragkit import PDFRAG

rag = PDFRAG("data/sample.pdf")
result = rag.ask("What is memory?", return_sources=True)

print(result["answer"])
print(result["sources"])

Example shape:

{
    "answer": "Memory in LangChain stores previous conversation turns...",
    "sources": [
        {
            "content": "Memory in chat applications is created by storing earlier conversation turns...",
            "page": 2,
            "source": "data/sample.pdf",
            "metadata": {
                "page": 2,
                "source": "data/sample.pdf"
            }
        }
    ]
}

---

ask() vs chat()

Method	Purpose
ask()	Stateless document Q&A
chat()	History-aware conversational interaction

Use ask() when you want a direct answer from the document.

Use chat() when you want:

• follow-up questions
• translation of the previous answer
• history-based conversation

---

LLM Providers

Default: Sarvam

from ragkit import PDFRAG

rag = PDFRAG("file.pdf")

OpenAI

from ragkit import PDFRAG

rag = PDFRAG(
    "file.pdf",
    llm_provider="openai",
    llm_config={
        "model": "gpt-4o-mini",
        "temperature": 0.1,
    },
)

Claude

from ragkit import PDFRAG

rag = PDFRAG(
    "file.pdf",
    llm_provider="claude",
    llm_config={
        "model": "claude-3-5-haiku-latest",
        "temperature": 0.2,
    },
)

Custom LLM

from langchain_openai import ChatOpenAI
from ragkit import PDFRAG

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
rag = PDFRAG("file.pdf", llm=llm)

---

Configuration

from ragkit import PDFRAG, RAGConfig

config = RAGConfig(
    chunk_size=800,
    chunk_overlap=150,
    top_k=5,
    use_multi_query=True,
    enable_query_rewrite=True,
)

rag = PDFRAG("file.pdf", config=config)

Configurable options currently include:

• persist_directory
• chunk_size
• chunk_overlap
• top_k
• use_multi_query
• enable_query_rewrite
• collection_name
• verbose
• llm_provider
• llm_model
• llm_temperature
• llm_kwargs

---

Add More Documents

rag.add_documents("data/another.pdf")

---

Reset Chat

rag.reset_chat("user1")

---

Project Structure

rag-kit/
├── .env.example
├── .gitignore
├── README.md
├── pyproject.toml
├── examples/
├── data/
├── src/
│   └── ragkit/
└── third_party/

---

Do You Need requirements.txt?

Not necessarily.

For modern Python packaging, pyproject.toml is enough and should be the main source of dependencies.

Use requirements.txt only if you want one of these:

• easier local setup for teammates
• pinned development environment
• quick install for people who do not use packaging workflows

Recommendation

Keep:

• pyproject.toml as the main dependency file

Optional:

• requirements-dev.txt for local development and testing

Example requirements-dev.txt:

pytest
black
ruff
build
twine

If you want, you can also generate a plain requirements.txt, but it should not replace pyproject.toml.

---

Current Limitations

• Primarily optimized for PDF-based RAG
• Sarvam support may depend on vendored or local integration setup
• No streaming support yet
• No FastAPI server or UI layer yet
• Agent support is planned, but not included in the current public API

---

Roadmap

• Better source citations
• Improved multi-file indexing isolation
• Streaming responses
• FastAPI server mode
• Playground / UI
• Agent support via ragkit.agent

---

Examples

Check the examples/ folder for runnable examples such as:

• basic_ask.py
• chat_example.py
• provider_openai.py

---

License

MIT License

---

Version

Current version: 0.1.0-beta

APIs may evolve in future releases.