Nyaya Dhwani

Multilingual legal information assistant for Indian law — ask questions about the Bharatiya Nyaya Sanhita (BNS), Indian Penal Code (IPC), and their mappings in 13 languages.

Built on Databricks Free Edition with FAISS RAG, Llama Maverick (AI Gateway), and Sarvam AI (translation, speech-to-text, text-to-speech). Deployed as a Databricks App via Gradio.

Not legal advice. General information only — consult a qualified lawyer for your specific situation.

How it works

Question (any of 13 languages)
  → Sarvam Mayura translates to English
  → FAISS semantic search (900+ legal text chunks)
  → Databricks Llama Maverick generates answer with citations
  → Sarvam Mayura translates back to selected language
  → Bilingual response (selected language + English) + sources

Supported languages: English, Hindi, Bengali, Kannada, Tamil, Telugu, Malayalam, Marathi, Gujarati, Odia, Punjabi, Assamese, Urdu.

Voice support: microphone input via Sarvam Saaras STT, answer read aloud via Sarvam Bulbul TTS.

Documentation

Document	Audience	What it covers
App User Guide	End users	How to use the app — language selection, asking questions, understanding responses, voice features
Developer Guide	Developers	Deploying the app, secrets/auth, Gradio + Databricks Apps pitfalls, translation pipeline, dependency pins
UI Design	Designers / developers	UI/UX spec and Sarvam pipeline design
Playground to App	Developers	Mapping Playground "Get code" to app env vars
Workspace Setup	Admins	Databricks secret scopes, GitHub Repos, key rotation
Benchmark Evaluation	Developers	RAG quality evaluation with BhashaBench-Legal + internal test questions
Plan	Team	Product plan and architecture decisions

Quick start

For users

Open the app URL → select language → ask a question. See App User Guide.

For developers

# 1. Authenticate
databricks auth login --host https://dbc-6651e87a-25a5.cloud.databricks.com --profile free-aws
export DATABRICKS_CONFIG_PROFILE=free-aws

# 2. Store secrets
databricks secrets create-scope nyaya-dhwani
databricks secrets put-secret nyaya-dhwani sarvam_api_key
databricks secrets put-secret nyaya-dhwani hf_token          # HuggingFace (for benchmark dataset)

# 3. Run notebooks (on a Databricks cluster)
#    - notebooks/india_legal_policy_ingest.ipynb  (ingest → legal_rag_corpus table)
#    - notebooks/build_rag_index.ipynb            (FAISS index → UC Volume)

# 4. Deploy the app
#    Compute → Apps → Create → connect this Git repo → Deploy

# 5. Grant service principal permissions
#    - CAN_QUERY on AI Gateway endpoint
#    - READ on UC Volume main.india_legal.legal_files
#    - READ on secret scope nyaya-dhwani

Full details: Developer Guide.

Local development

pip install -e ".[dev,rag,rag_embed,app]"
cp .env.example .env   # fill in values
export $(grep -v '^#' .env | xargs)
python app/main.py

Repository layout

Path	Purpose
app/main.py	Gradio app (RAG + LLM + Sarvam multilingual)
app.yaml	Databricks Apps entry point + env config
src/nyaya_dhwani/	Python package: embedder, retrieval, llm_client, sarvam_client
notebooks/	Data ingestion + FAISS index build
requirements.txt	Databricks Apps pip install
tests/	pytest suite
docs/	All documentation (see table above)

Testing

pip install -e ".[dev,rag]"
pytest tests/ -v

Technology stack

Component	Technology
LLM	Databricks Llama 4 Maverick (AI Gateway)
Embeddings	sentence-transformers/all-MiniLM-L6-v2
Vector search	FAISS (IndexFlatIP, cosine similarity)
Translation	Sarvam Mayura
Speech-to-text	Sarvam Saaras
Text-to-speech	Sarvam Bulbul
App framework	Gradio 4.44 on Databricks Apps
Data platform	Databricks (Unity Catalog, Volumes, Apps)