README.md

---
title: OSH Agentic RAG
emoji: 🤖
colorFrom: purple
colorTo: blue
sdk: docker
pinned: false
tags:
- ml-intern
---

# 🤖 OSH Agentic RAG

An **agentic** RAG system that uses tool-calling to answer legal questions about Kenyan OSH law.

## How it works

Unlike traditional RAG (retrieve → generate), this agent **decides what to do**:

1. **Receives question** → analyzes what information is needed
2. **Calls tools** → `search_documents` or `search_specific_act`
3. **Reviews results** → decides if it has enough context
4. **Generates answer** → with citations to specific Acts/sections

## Architecture

| Component | Role | Size |
|-----------|------|------|
| **Qwen3 1.7B** (Q4_K_M) | Agent brain with tool-calling | 1.1 GB |
| **BGE-small-en-v1.5** | Semantic search embeddings | 130 MB |
| **FAISS** | Vector similarity search | — |

## Speed

Expected ~**10-20s per response** on free HF CPU (vs 75-90s for the non-agentic version).

## API

```python
from gradio_client import Client
client = Client("Rofati/osh-agentic-rag")
result = client.predict("What are noise exposure limits?", api_name="/ask")
```