Pravah v2

Pravah is a search-first chat app that combines web results and agent tools. The name means "flow" in Sanskrit.

What it does

• Web search with source-backed answers
• Page fetching and long-document chunking
• Multi-provider LLM support via LiteLLM
• Persistent chat history in DuckDB
• Streaming responses and tool visibility

Quick start

Prerequisites

• Python 3.11+
• uv (recommended) or pip

Install

git clone https://github.com/jayshah5696/pravah.git
cd pravah

uv sync

Configure

Create a .env file in the repo root:

TVLY_API_KEY=your_tavily_api_key

OPENAI_API_KEY=your_openai_api_key
ANTHROPIC_API_KEY=your_anthropic_api_key
GEMINI_API_KEY=your_gemini_api_key
GROQ_API_KEY=your_groq_api_key
DEEPSEEK_API_KEY=your_deepseek_api_key

LANGCHAIN_API_KEY=your_langsmith_api_key
LANGCHAIN_PROJECT=pravah

If you want to customize the model list or UI, edit config.yaml.

Run

uv run streamlit run app.py

Open http://localhost:8501

Project structure

pravah/
├── app.py                 # Streamlit UI
├── config.yaml            # Model and UI configuration
├── pravah/
│   ├── agent.py           # LangGraph agent
│   ├── tools.py           # Tool layer (search, fetch, calculate)
│   ├── prompts.py         # System prompt builder
│   ├── history.py         # DuckDB history storage
│   ├── search.py          # Search providers
│   ├── retrieval.py       # Chunking and reranking
│   └── llm.py             # LLM helpers
├── scripts/
│   └── eval.py            # Evaluation runner
├── tests/
│   └── eval_set.csv       # Eval dataset
├── .streamlit/
│   └── config.toml        # Theme
└── docs/
    └── EVAL_GUIDE.md      # Evaluation notes

Models and providers

The UI reads available models from config.yaml. Keep that file as the source of truth for the UI list. If you use a custom model string, enter it in the sidebar.

Agent tools

Tool	Purpose
web_search	Web search via Tavily
fetch_page	Fetch and extract page text
read_page_chunk	Navigate long pages
search_memory	Search previously fetched content
calculate	Safe math evaluation

Evaluation

uv run python scripts/eval.py
uv run python scripts/eval.py --limit 10
uv run python scripts/eval.py --verbose

Results are written to tests/eval_results.csv and tests/traces/.

Docker

docker build -t pravah .
docker run -p 8501:8501 --env-file .env pravah

Development

uv sync --group dev
uv run pytest
uv run ruff check .
uv run ruff format .

Contributing

Open an issue or PR with a focused change and a short description of how you tested it.

License

MIT. See LICENSE.