Polyglot RAG Lab

A multilingual Basic + Advanced RAG showcase using notebooks, Gradio, vector databases, document parsing, chunking, embeddings, and retrieval-based answering.

This project demonstrates how a retrieval-augmented generation pipeline can answer questions from a small knowledge base and handle multilingual queries such as English and German.

Project highlights

• Basic RAG workflow with vector database initialization
• Advanced RAG answer generation workflow
• Gradio interface for asking questions over a local knowledge base
• Multilingual retrieval demonstration with German and English queries
• Sample data prepared for parsing, chunking, embedding, and vector storage
• Documentation-friendly structure for reports, notebooks, and portfolio review

Demo proof

Multimodal / document RAG pipeline

Basic RAG answer in English

Basic RAG answer in German

Repository structure

polyglot-rag-lab/
├── data/
│   ├── raw/
│   │   ├── markdown/
│   │   │   └── german_vegan_food_guide.md
│   │   ├── text/
│   │   │   ├── italian_recipe_notes.txt
│   │   │   └── audio_transcript_sample.txt
│   │   ├── tables/
│   │   │   └── european_vegan_recipes.csv
│   │   └── multimodal/
│   │       └── image_caption_samples.csv
│   ├── processed/
│   │   └── chunks_preview.jsonl
│   └── vector_store/
│       └── .gitkeep
├── docs/
│   └── pipeline.md
├── notebooks/
│   └── README.md
├── screenshots/
│   ├── multimodal-rag-pipeline.svg
│   ├── basic-rag-english-demo.svg
│   └── basic-rag-german-demo.svg
├── .gitignore
├── requirements.txt
└── README.md

End-to-end RAG pipeline

The project follows a practical RAG workflow:

Raw files
  ↓
Parsing / loading
  ↓
Text normalization
  ↓
Chunking
  ↓
Embedding generation
  ↓
Vector database storage
  ↓
Retriever
  ↓
Prompt construction with retrieved context
  ↓
LLM answer generation
  ↓
Structured response for reports, notebooks, or documentation

1. Raw data layer

The data/raw/ folder contains sample inputs that represent the kind of content a RAG system may need to parse.

Folder	Purpose	Example
markdown/	Long-form structured documents	German vegan food guide
text/	Plain text, transcripts, notes	Audio transcript and recipe notes
tables/	CSV-style structured knowledge	European vegan recipe table
multimodal/	Captions or extracted descriptions from images/videos	Image caption samples

In a real multimodal RAG system, images, audio, PDFs, video, CSV, and Markdown files are first converted into text or structured metadata before embedding.

2. Parsing layer

Typical parsing tools may include:

• SimpleDirectoryReader from LlamaIndex for text, Markdown, PDF, CSV, and other document formats
• OCR for scanned images or screenshots
• Image captioning models for visual content
• Speech-to-text models for audio
• Video frame extraction plus captioning/transcription for video

The output of parsing should be normalized text with useful metadata, for example:

{
  "source": "data/raw/markdown/german_vegan_food_guide.md",
  "modality": "markdown",
  "language": "en",
  "topic": "German vegan food",
  "text": "Kartoffelsalat is a common German dish that can be made vegan..."
}

3. Chunking layer

Large documents are split into smaller chunks before embedding. Good chunks should be:

• Small enough to retrieve precisely
• Large enough to preserve meaning
• Stored with source metadata
• Overlapped slightly to avoid losing context across boundaries

Example chunking settings:

chunk_size = 512
chunk_overlap = 50

4. Embedding layer

Each chunk is converted into a numerical vector using an embedding model. These vectors allow semantic search, so the system can retrieve relevant content even when the user asks in different words or a different language.

Example:

Question: Was sind typische deutsche Gerichte?
Retrieved chunk: German dishes include Kartoffelsalat, Sauerkraut, Bratwurst, Schnitzel, and regional stews...

5. Vector database layer

The vector database stores:

• Chunk text
• Embeddings
• Metadata
• Source paths
• Optional language and modality tags

The data/vector_store/ folder is included as a placeholder. Generated vector database files are usually ignored from Git because they can become large and environment-specific.

6. Retrieval layer

When the user asks a question, the system:

1. Embeds the question
2. Searches the vector database
3. Retrieves the most relevant chunks
4. Builds a context-aware prompt
5. Sends the prompt to the language model

7. Answer generation layer

The model answers using retrieved context, rather than relying only on its internal training data. This makes the system more grounded, auditable, and useful for domain-specific knowledge bases.

Example multilingual behavior:

German question:
Was sind typische deutsche Gerichte? Nenne mir bitte auch ein Rezept vegan.

Expected RAG behavior:
Retrieve German food chunks, identify typical dishes, then generate a vegan recipe such as vegan Kartoffelsalat.

Suggested GitHub description

A multilingual Basic and Advanced RAG showcase using vector databases, Gradio apps, sample data, and notebook-based workflows.

Suggested topics

rag, advanced-rag, basic-rag, multilingual-rag, gradio, vector-database, llamaindex, embeddings, ai-engineering, notebooks

Next improvements

• Add the original assignment notebooks into notebooks/
• Add advanced RAG screenshots showing retrieved context and improved answer quality
• Add source citation display inside the Gradio app
• Add hybrid retrieval using vector search plus keyword search
• Add reranking and evaluation metrics