You can deploy this model and start earning money today!

tanuj437/
bhagavat-gita-rag

Name: bhagavat-gita-rag
Availability: InStock

safetensors

text-generation

google/muril-base-cased

finetune

mit

Views

No views yet

Bhagavad Gita RAG

Bhagavad Gita RAG is a release-ready dual-retriever Retrieval-Augmented Generation project for grounded question answering over the Bhagavad Gita. It combines:

SanskritBERT for Sanskrit-first semantic retrieval
MuRIL for multilingual retrieval across English, Hindi, and Devanagari queries
Qwen 2.5 Instruct for answer generation

The repository is structured so the main way to run the project is through rag-run.py, while the notebook remains available for interactive exploration.

What This Repository Includes

A clean Python package in src/gita_rag/
A single launcher entry point: rag-run.py
Prebuilt retriever assets in dual_retriever_models_clean/
Corpus and train/dev/test JSONL files in dual_retriever_data/
A lightweight notebook wrapper in rag-run.ipynb
An archived training notebook in dual-model-finetuning.ipynb

Main Workflow

Run the project in one of two ways:

python rag-run.py --ui

Or ask one question directly:

python rag-run.py --question "Why is Arjuna overcome with sorrow before the war?"

Installation

Create and activate a Python environment.
Install dependencies:

pip install -r requirements.txt

Make sure these folders stay in the repository root:

dual_retriever_data/
dual_retriever_models_clean/

Project Layout

text

bhagavad-gita-rag/
|-- rag-run.py                     # Main launcher
|-- rag-run.ipynb                 # Notebook wrapper around the packaged pipeline
|-- dual-model-finetuning.ipynb   # Archived training notebook
|-- src/gita_rag/
|   |-- config.py                 # Paths and runtime configuration
|   |-- retrieval.py              # Dense retriever loading and search
|   |-- generation.py             # vLLM / Transformers generation backends
|   |-- pipeline.py               # End-to-end RAG orchestration
|   |-- text.py                   # JSONL loading and text cleanup helpers
|   `-- ui.py                     # Gradio interface
|-- dual_retriever_data/          # Corpus + retrieval train/dev/test files
|-- dual_retriever_models_clean/  # Fine-tuned retriever checkpoints + FAISS indices
|-- requirements.txt
`-- pyproject.toml

How Retrieval Works

For every user question:

The query is embedded with both retrievers.
Each retriever searches its FAISS index.
Results are fused into a final ranked set.
The top verses are formatted as context.
Qwen generates an answer constrained to the retrieved verses.

The current fusion weights favor MuRIL because the supplied evaluation file shows it is the stronger retriever on the provided multilingual benchmark, while SanskritBERT is still preserved as a Sanskrit-specialized component.

Model Notes

SanskritBERT is kept as the Sanskrit-aware retriever.
MuRIL is kept as the multilingual retriever.
Qwen/Qwen2.5-7B-Instruct is the default generation model.
If vLLM is unavailable, the code falls back to a standard transformers backend so the repository still runs in more environments.

Important Notes

The packaged loader repairs the visible text mojibake found in the shipped JSONL files at runtime.
vLLM is typically best on Linux with CUDA. On Windows, the project will usually fall back to transformers.
The training notebook is kept as a reference artifact, not as the primary way to understand or run the project.

Evaluation Snapshot

The repository includes dual_retriever_models_clean/comparison_clean.json, which shows the current benchmark summary:

MuRIL is the higher-performing retriever on the provided evaluation split.
SanskritBERT remains included for Sanskrit-specific retrieval behavior.

License

This project is released under the MIT License.

tanuj437/bhagavat-gita-rag