Agentic Search

Multi-agent research search — Ask a question, and a team of AI agents plans searches, fetches results, scrapes pages, writes cited reports, and iterates until the answer is solid.

---

What It Does

Agentic Search turns a single research question into a full research pipeline. Instead of one generic search, specialized agents:

• Plan the best search strategy and query
• Search the web via Google Serper
• Select the most relevant result from the SERP
• Scrape the chosen page for content
• Report an answer with citations
• Review the report and decide: done, or loop back to improve

You get a cited, reviewed answer that can go through multiple refinement loops (new searches, different sources, or better writing) until quality is acceptable.

---

Features

• Multi-agent workflow — Planner, Selector, Reporter, Reviewer, and Router agents with clear roles
• Live web search — Google Serper API for real search results
• Page scraping — Fetches and parses the selected URL (BeautifulSoup), with error handling and content limits
• Review loop — Reviewer checks reports; Router sends work back to Planner, Selector, or Reporter as needed
• Structured outputs — Agents use JSON schemas (Groq JSON mode) for reliable parsing and routing
• Streamlit UI — Chat-style interface; you ask a question and see each agent’s output as it runs
• Configurable LLM — Built for Groq (e.g. llama3-8b-8192); API keys via config/config.yaml

---

Workflow

End-to-end flow from your question to the final report:

1. Planner — Turns your question into a search term and strategy.
2. Serper — Runs the search and returns organic results (title, link, snippet).
3. Selector — Picks one result and returns its URL.
4. Scraper — Downloads and parses that page (text, length limit, error handling).
5. Reporter — Writes an answer with citations from the scraped content.
6. Reviewer — Evaluates the report (comprehensiveness, citations, relevance).
7. Router — Chooses next step: final_report (done), or loop back to planner / selector / reporter.
8. Final report — When the Router decides done, the last report is shown as the answer.

---

Prerequisites

• Python 3.x
• API keys
  • Groq — for the LLM (e.g. Llama 3)
  • Serper — for Google search results

---

Setup

1. Clone and enter the repo

   git clone https://github.com/omkarjoshi9/Agentic_Search.git
   cd "Agentic Search"

2. Create a virtual environment (recommended)

   python -m venv venv
   # Windows
   venv\Scripts\activate
   # macOS/Linux
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Configure API keys

   Edit config/config.yaml and set your keys (do not commit real keys):

   groq_api_key: "your-groq-api-key"
   serper_api_key: "your-serper-api-key"
   openai_api_key: ""   # optional, for future use

---

Usage

From the project root:

streamlit run app/app.py

Then:

1. Open the URL shown in the terminal (usually http://localhost:8501).
2. Type your research question in the chat input (e.g. "Why is the sky blue?").
3. Watch the agents run: Planner → Serper → Selector → Scraper → Reporter → Reviewer → Router, with optional loops.
4. Read the Final Report when the workflow finishes.

---

Project Structure

Agentic Search/
├── app/
│   └── app.py              # Streamlit entry point, runs the workflow
├── agent_graph/
│   └── graph.py            # LangGraph definition (nodes, edges, router)
├── agents/
│   └── agents.py           # Planner, Selector, Reporter, Reviewer, Router, FinalReport, End
├── config/
│   └── config.yaml         # API keys (groq, serper, openai)
├── models/
│   └── groq_models.py      # Groq LLM wrappers (JSON and plain)
├── prompts/
│   └── prompts.py          # System prompts and JSON schemas per agent
├── states/
│   └── state.py            # AgentGraphState and state helpers
├── tools/
│   ├── google_serper.py    # Serper API client
│   └── basic_scraper.py    # HTTP fetch + BeautifulSoup scraping
├── utils/
│   ├── steamlit.py         # Streamlit UI (input, message queue, display)
│   ├── message_queue.py    # Queue for agent messages to the UI
│   └── helper_functions.py # Config loading, datetime, content checks
├── images/
│   └── image.png           # Workflow diagram
├── requirements.txt
└── readme.md

---

Tech Stack

Layer	Technology
Orchestration	LangGraph (state graph, conditional routing)
LLM	Groq (e.g. Llama 3 8B), JSON and text modes
Search	Serper (Google search API)
Scraping	requests + BeautifulSoup
UI	Streamlit
Config	YAML (config/config.yaml)

---

Configuration

• LLM — In app/app.py: server, model, and optional model_endpoint.
• Loop limit — In app/app.py: limit = {"recursion_limit": iterations} (e.g. 40) to cap review loops.
• API keys — All in config/config.yaml; loaded into env by load_config() in the tools/models that need them.

---

License

See the repository license file for terms of use.