Agentic Multi-Tool Research Assistant

A research-oriented assistant that can answer from uploaded PDFs or deliberately invoke external tools for current and general information.

Distinct use case

This repository demonstrates tool orchestration and source routing. It is different from the NVIDIA knowledge-base project, which uses only indexed documents and hybrid retrieval.

Question type	Route
Question grounded in uploaded reports	PDF retrieval chain
Current information or news	DuckDuckGo
Scientific paper lookup	arXiv
General background	Wikipedia

A transparent routing policy sends time-sensitive questions to web tools even when PDFs are loaded.

Architecture

flowchart LR
    A["User question"] --> B["Source router"]
    B --> C["PDF retrieval"]
    B --> D["Search tool"]
    B --> E["arXiv tool"]
    B --> F["Wikipedia tool"]
    C --> G["Answer with conversation memory"]
    D --> G
    E --> G
    F --> G

Loading

Run

git clone https://github.com/sandeep848/agentic-multi-tool-chatbot.git
cd agentic-multi-tool-chatbot
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
streamlit run main.py

Test

pytest

Key engineering decisions

• uploaded files are fingerprinted before index reuse
• PDF answers are constrained to retrieved document context
• time-sensitive cues force a web-tool route
• conversation history is isolated by session
• the active model and retrieval preference remain user-controlled

Limitations

External search results are untrusted inputs and should be verified. The application is a research assistant, not an autonomous decision maker.