MULIA-Multilingual Legal Information Assistant

Inspiration

I was inspired by a simple gap: legal information is often available, but not accessible in the language, tone, or speed that people need in real moments. The goal became clear: build a multilingual legal assistant that feels practical, trustworthy, and fast for everyday users.

What it does

How we built it

I designed the project as a Retrieval-Augmented Generation (RAG) assistant with a clean local app experience and Databricks-native deployment path.

Built ingestion pipelines for legal datasets (constitution, criminal sections, government schemes). Normalized and chunked text for retrieval quality. Added retrieval backends (Vector Search primary, FAISS-compatible path where supported). Integrated an OpenAI-compatible LLM call path for legal Q&A. Added multilingual and voice-adjacent hooks for broader accessibility. Wrapped everything in a Gradio interface with a calm, readable UI theme. The basic retrieval objective I used was balancing relevance and trust:

Challenges we ran into

Package compatibility issues on newer Python versions. Backend auth differences between local and cloud environments. Endpoint path mismatches causing authorization failures. Retrieval backend differences (FAISS vs managed vector search). Keeping UI readability high while expanding features.

Accomplishments that we're proud of

What we learned

Reliability matters more than cleverness in legal UX. Small environment mismatches can break production behavior. Dependency compatibility is a first-class engineering task. Good prompts help, but good data preparation helps more. Clear fallback paths are essential for real-world resilience.

Built With

• databricks
• llama
• python
• sarvam