Clickless AI

One drop zone. No mode-picking. Drop a file (PDF, text, image) or ask a question that needs live web info, and a router automatically decides which specialized tool should handle it — you never have to tell it what kind of task it is.

How it works

File / Query
     │
     ▼
Type Detector ──► Loader (PDF / text / image) ──► Chunker ──► FAISS index
     │
     ▼
Router (LLM) ──► decides: tool + retrieval query + k
     │
     ├── chat_summarizer  → summarizes a chat/conversation
     ├── pdf_digester      → explains / answers questions about a document
     ├── draft_reducer     → tightens/shortens a piece of writing
     ├── image_tool        → describes/analyzes an image
     └── web_search        → answers using live Tavily search results
     │
     ▼
   Answer

Built with LangGraph as a state machine — the router node picks the next node with a conditional edge, so adding a new tool means adding one node + one edge, not new if/else branches.

Multi-provider fallback

Both the router and the main answering LLM run through a fallback chain (Groq → Gemini), so a single provider outage or rate limit doesn't take the whole pipeline down mid-demo.

Project structure

clickless_ai/
├── main.py          # graph definition: router + tool nodes + fallback LLM chains
├── file_loader.py   # PDF / text / image loaders (vision model for images)
├── vectorstore.py   # chunking + FAISS indexing + retrieval
├── app.py           # Streamlit UI (drag-and-drop upload)
├── requirements.txt
├── .env.example
└── .gitignore

Setup

git clone https://github.com/Dwarkesh-code/clickless_ai.git
cd clickless_ai
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env       # then fill in your API keys

Run

Streamlit UI (recommended):

streamlit run app.py

Drag and drop a file, type your question, and the router handles the rest.

CLI:

python main.py

Environment variables

See .env.example. You need API keys for Groq, Google (Gemini), and Tavily. The image tool additionally uses an NVIDIA vision endpoint as its primary model with Gemini vision models as fallback.

Status / known limitations

• Website ingestion (detect_input_type returns "website") is not implemented yet — a URL input will raise NotImplementedError until a website loader is added to file_loader.py.
• The FAISS index is wiped on start and on exit — there is no persistence across runs by design (each session starts clean).
• No automated tests yet.

Roadmap

• Website loader (requests + BeautifulSoup)
• Persistent/optional vectorstore across sessions
• Streaming answers in the Streamlit UI
• More tool types