AI Multi-Research Agent

A four-stage multi-agent research pipeline built with LangChain and Streamlit. A search agent, a reader/scraper agent, a writer chain, and a critic chain work together to turn a topic into a structured, cited research report with a critique.

Features

• Search agent — tool-calling agent (create_agent) that queries DuckDuckGo for titles, URLs, and snippets
• Reader agent — tool-calling agent that scrapes and cleans a chosen page (requests + BeautifulSoup), truncated to ~3000 characters
• Writer chain — prompt chain that drafts a structured report (Introduction / Key Findings / Conclusion / Sources)
• Critic chain — prompt chain that scores the report out of 10 with strengths, areas to improve, and a one-line verdict
• Both a Streamlit web UI and a scripted CLI runner

Tech Stack

• Orchestration: LangChain (langchain, langchain-core, langchain-community)
• LLM: OpenRouter, via langchain-openrouter's ChatOpenRouter — default model openai/gpt-oss-20b:free
• Search: duckduckgo_search, with fallback handling across its ddg function and DDGS class APIs (the package's interface has changed across versions)
• Scraping: requests + beautifulsoup4 (+ lxml/html5lib parsers)
• UI: Streamlit

Pipeline

Topic
  │
  ▼
Search Agent   → web_search (DuckDuckGo)                → titles, URLs, snippets
  │
  ▼
Reader Agent   → scrape_url (requests + BeautifulSoup)   → cleaned page text (~3000 chars)
  │
  ▼
Writer Chain   → structured report (Introduction / Key Findings / Conclusion / Sources)
  │
  ▼
Critic Chain   → scored review (Strengths / Areas to Improve / Verdict)

All four stages run on the same LLM (openai/gpt-oss-20b:free via OpenRouter, temperature=0, max_tokens=1000 by default). Only the Search and Reader stages are tool-calling agents (create_agent); the Writer and Critic stages are plain prompt chains (ChatPromptTemplate | llm | StrOutputParser).

Installation

git clone https://github.com/chmodgaurav/AI-Multi-Research-Agent.git
cd AI-Multi-Research-Agent
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

Configuration

Create a .env file in the project root:

OPENROUTER_API_KEY=your_openrouter_api_key

agents.py initializes ChatOpenRouter(model="openai/gpt-oss-20b:free", ...), which reads OPENROUTER_API_KEY from the environment.

Usage

Web UI

streamlit run streamlit_app.py

Open http://localhost:8501, enter a research topic, and the app runs through search → read → write → critique.

CLI

python pipeline.py

Prompts for a topic in the terminal, runs the same four-stage pipeline, and prints each stage's output as it completes.

Project Structure

File	Purpose
streamlit_app.py	Web UI entry point (Streamlit)
pipeline.py	CLI runner for the full pipeline (run_research_pipeline)
agents.py	Agent/chain definitions and LLM configuration
tools.py	web_search and scrape_url tool implementations
requirements.txt	Python dependencies

Implementation Notes

• Search (tools.web_search): wraps duckduckgo_search, trying the module's ddg function first, then its DDGS class, then a last-resort submodule lookup — this covers both older and newer versions of the package.
• Scraper (tools.scrape_url): fetches and cleans page text, truncated to ~3000 characters to keep prompts within context limits.
• Model: change the model, temperature, or token limit in agents.py (llm = ChatOpenRouter(...)).
• Writer/Critic: both are plain prompt chains, not tool-calling agents — they receive text (topic + combined research, or the drafted report) and return text.

Troubleshooting

• duckduckgo_search import errors: the package's API has changed across versions; tools.py already handles both the legacy ddg function and the newer DDGS class, but ensure the package is installed (pip install duckduckgo_search).
• LLM authentication errors: confirm OPENROUTER_API_KEY is set and valid, and that the account has credit/access for the configured model.
• Missing packages in Streamlit: re-run pip install -r requirements.txt inside the activated virtual environment.

Future Improvements

• Make the LLM model configurable via environment variable instead of hardcoding it in agents.py
• Add automated tests (tests/ directory does not currently exist)
• Support multi-source reading (currently the reader agent scrapes a single chosen URL per run)

Screenshots

Contributing

Issues and pull requests are welcome.

License

See repository for license details.