waseek/multi-agent-tool-calling
1
Research Intelligence Assistant
A production-grade AI research platform built with Streamlit, LangGraph, and LangChain. It combines a multi-provider LLM stack, a 13-tool ReAct agent, PDF paper analysis with semantic search, user authentication, and token-by-token streaming into a single deployable app.
    
Live Demo: Hugging Face Spaces
Architecture
+----------------------------------------------------------+
| Streamlit UI |
| Hero header | Sidebar | Chat | PDF uploader |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Auth Gate |
| SQLite + bcrypt | 20 chats/day limit | Admin bypass|
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| LangGraph ReAct Graph |
| |
| START |
| | |
| v |
| conversation_manager |
| - trims history to last N messages |
| - builds rolling conversation summary |
| | |
| v |
| context_llm <-----------------------+ |
| - selects model by task type | |
| - rebuilds system prompt per turn | |
| - binds all 13 tools | |
| | | |
| +-- tool call? --> enhanced_tools -+ |
| | (runs tool, returns ToolMessage) |
| | |
| +-- final answer --> END (streamed token by token) |
+----------------------------------------------------------+
| |
v v
+-------------------+ +-------------------+
| LLM Providers | | Tool Layer |
| | | |
| 1. Anthropic | | Academic Search |
| Claude Opus | | - ArXiv |
| Claude Sonnet | | - PubMed |
| Claude Haiku | | - Semantic Scholar|
| | | - OpenAlex |
| 2. OpenAI | | - Find Related |
| GPT-4o | | |
| GPT-4o Mini | | Web & General |
| | | - DuckDuckGo |
| 3. Groq | | - Tavily |
| Llama 3.3 70B | | - Wikipedia |
| Llama 4 Scout | | |
| Llama 3.1 8B | | PDF + Utilities |
| | | - pdf_search |
| Auto fallback | | - Calculator |
| on rate limits | | - Code Analyzer |
+-------------------+ | - Weather Info |
| - File Generator |
+-------------------+
|
v
+-------------------+
| Data Layer |
| |
| ChromaDB |
| (vector store) |
| sentence- |
| transformers |
| embeddings |
| |
| SQLite |
| (users + usage) |
+-------------------+Features
- Multi-provider LLM - Anthropic Claude Opus/Sonnet/Haiku, OpenAI GPT-4o/Mini, Groq Llama 3.3/4/3.1 with automatic failover on rate limits
- Live model selector - switch between any available model mid-conversation from the sidebar without losing history
- 13 research tools - academic databases, web search, PDF RAG, calculator, code analysis, and more
- PDF paper analysis - upload research papers, ask questions, get page-level citations; powered by ChromaDB and sentence-transformers
- User auth - register/login with bcrypt-hashed passwords, 20 chats/day limit, admin bypass
- Token-by-token streaming - tool progress indicators appear inline before the final answer streams
- Tool badges - colored pill badges show which tools were used in each response
- Copy button - small icon copies the full response to clipboard (works inside Streamlit's iframe sandbox)
- Conditional references - structured citations appended only when tools were used or factual claims were made
Tools
Academic Research
Web and General
Utilities
LLM Provider Chain
The system auto-detects available API keys and builds the fallback chain at startup:
If ANTHROPIC_API_KEY is set:
1. claude-opus-4-8 (primary - most capable, best research reasoning)
2. claude-sonnet-4-6 (secondary - fast, excellent quality)
3. gpt-4o-mini (fallback - if OPENAI_API_KEY also set)
4. llama-3.3-70b (fallback - if GROQ_API_KEY also set)
If only OPENAI_API_KEY is set:
1. gpt-4o (primary)
2. gpt-4o-mini (secondary)
3. llama-3.3-70b (fallback - if GROQ_API_KEY also set)
If only GROQ_API_KEY is set:
1. llama-3.3-70b-versatile (primary)
2. meta-llama/llama-4-scout-17b-16e-instruct (secondary)
3. llama-3.1-8b-instant (fallback)Failed models enter a 60-second cooldown before being retried. Task routing selects temperature automatically: 0.0 for math and code, 0.1 for research and analysis, 0.7 for creative tasks.
Project Structure
Multi-tool Research Bot Assistant/
├── app.py # Streamlit UI, LangGraph graph assembly, streaming
├── requirements.txt
├── .streamlit/
│ ├── config.toml # Theme, server settings
│ └── secrets.toml # API keys for cloud deployment (gitignored)
├── data/ # Auto-created at runtime (gitignored)
│ ├── auth.db # SQLite: users and daily usage
│ └── chroma_db/ # ChromaDB: PDF vector embeddings
└── src/
├── __init__.py
├── models.py # EnhancedLLM (multi-provider failover), ConversationState
├── nodes.py # LangGraph nodes: context_llm, enhanced_tools, task routing
├── tools.py # All 13 tool definitions, contextvars for user/PDF context
├── conversation.py # History trimming and rolling conversation summary
├── auth.py # SQLite auth: register, login, daily limit enforcement
└── rag.py # PDFProcessor (PyMuPDF) + ResearchVectorStore (ChromaDB)Getting Started
Requirements: Python 3.11+, at least one API key (Anthropic, OpenAI, or Groq)
git clone https://github.com/Waseekk/multi-tool-research-bot.git
cd multi-tool-research-bot
python -m venv venv
venv\Scripts\activate # Windows
# source venv/bin/activate # macOS / Linux
pip install -r requirements.txtCreate a .env file (include whichever keys you have):
ANTHROPIC_API_KEY=sk-ant-... # Recommended - Claude Opus 4.8
OPENAI_API_KEY=sk-proj-... # Optional - GPT-4o fallback
GROQ_API_KEY=gsk_... # Optional - free tier fallback
TAVILY_API_KEY=tvly-... # Optional - enhanced web search
ADMIN_PASSWORD=your_admin_password # Password for the admin accountRun:
streamlit run app.pyThe first PDF upload downloads the embedding model (~420 MB, one-time only). Subsequent uploads are fast.
Deployment
Hugging Face Spaces
- Push to a Hugging Face Space repository
- Add API keys under Settings > Repository secrets
Streamlit Cloud
- Push to GitHub
- Connect at streamlit.io/cloud
- Add API keys under Settings > Secrets
Docker
docker build -t research-bot .
docker run -p 8501:8501 --env-file .env research-botTech Stack
License
MIT