Multi-Format RAG Assistant

A Retrieval-Augmented Generation (RAG) application built with FastAPI, LangChain, Google Gemini, and ChromaDB. The application allows users to upload PDF documents, generate embeddings, store them in a vector database, and ask natural language questions based on the uploaded content.

---

Features

• Upload PDF documents
• Validate uploaded files
• Extract text using PyPDFLoader
• Split documents into chunks using RecursiveCharacterTextSplitter
• Generate embeddings using Google Gemini
• Store document embeddings in ChromaDB
• Retrieve relevant document chunks using semantic search
• Generate context-aware answers using Gemini LLM
• Return source citations including:
  • Document Name
  • Page Number
  • Chunk ID
• Interactive API documentation using FastAPI Swagger

---

Tech Stack

• Python 3.13
• FastAPI
• LangChain
• Google Gemini API
• ChromaDB
• Pydantic Settings
• UV Package Manager

Installation

Clone the Repository

git clone https://github.com/rishurajgit/multi-format-rag.git

cd multi-format-rag

---

Install Dependencies

uv sync

---

Environment Variables

Create a .env file in the project root.

Example:

GEMINI_API_KEY=your_api_key

LLM_MODEL=your_Model_name

EMBEDDING_MODEL=your_embedding_model

CHUNK_SIZE=1000
CHUNK_OVERLAP=200

CHROMA_DB_PATH=./chroma_db
UPLOAD_DIR=uploads

---

Running the Application

uv run uvicorn main:app --reload

The server will start at:

http://127.0.0.1:8000

---

API Documentation

Swagger UI

http://127.0.0.1:8000/docs

ReDoc

http://127.0.0.1:8000/redoc

---

API Endpoints

Upload PDF

POST

/upload

Uploads a PDF document, extracts text, creates embeddings, and stores the document in ChromaDB.

Example Response

{
  "message": "PDF uploaded successfully",
  "document_name": "RAG Project 1.pdf",
  "total_pages": 3,
  "total_chunks": 6
}

---

Chat

POST

/chat

Request

{
  "question": "What is RAG?"
}

Example Response

{
  "answer": "RAG stands for Retrieval-Augmented Generation...",
  "sources": [
    {
      "document_name": "RAG Project 1.pdf",
      "page": 1,
      "chunk_id": 3
    }
  ]
}

---

Current Workflow

PDF Upload
      │
      ▼
Document Loader
      │
      ▼
Text Splitter
      │
      ▼
Gemini Embeddings
      │
      ▼
ChromaDB
      │
      ▼
Retriever
      │
      ▼
Prompt Template
      │
      ▼
Gemini LLM
      │
      ▼
Answer + Source Citation

---

Current Limitations

• Currently supports PDF documents only.
• Conversation history is not maintained.
• Retrieval always returns the top matching chunks.
• Multiple document format support (DOCX, TXT, Markdown) is planned.

---

Future Enhancements

• DOCX support
• TXT support
• Markdown support
• Similarity score threshold
• Streaming responses
• Conversation memory
• Multi-document collections
• Better citation formatting
• Docker

---

Author

Rishu Raj

GitHub: https://github.com/rishurajgit