Spaces:
Sleeping
Sleeping
metadata
license: mit
title: AutoScholar
sdk: docker
pinned: false
AutoScholar
AutoScholar is an autonomous multi-step research agent/assistant built with LangGraph. Given any research topic, it plans a search strategy, queries the web, reads and extracts content from sources, and formulates a structured, cited report — without any manual intervention. It uses a self-reflection loop to evaluate the quality of its output and automatically re-searches if the report doesn't fully answer the original question.
How It Works
When a user submits a research topic, the agent runs through a graph of nodes — each responsible for one stage of the research process:
- Planner — The LLM breaks the topic into 3-4 specific sub-questions worth searching for independently.
- Search — Each sub-question is sent to the Tavily Search API, which returns clean, LLM-ready results including URLs and content snippets.
- Reader — The top URLs are fetched and parsed with BeautifulSoup, extracting the main body text and stripping unwanted data.
- Synthesizer — All gathered content is assembled into a prompt and sent to the LLM, which writes a structured report with an introduction, key findings, and conclusion.
- Reflection — The LLM evaluates whether the report fully answers the original question. If not, the agent loops back to the Search node with refined queries. The loop is capped at 2 iterations to prevent infinite cycles.
Agent Graph
[Planner] → [Search] → [Reader] → [Synthesizer] → [Reflection]
↑ ↓
└──────────── re-search if needed ───────┘
↓
END
Project Structure
AutoScholar/
├── app/
│ ├── agent/
│ │ ├── nodes/
│ │ │ ├── planner.py
│ │ │ ├── search.py
│ │ │ ├── reader.py
│ │ │ ├── synthesizer.py
│ │ │ └── reflection.py
│ │ ├── graph.py
│ │ └── state.py
│ └── streamlit.py
└── README.md
Workflow
User Topic
↓
Planner (LLM)
↓
Tavily Web Search × N sub-questions
↓
BeautifulSoup Page Reader
↓
Synthesizer (LLM)
↓
Reflection (LLM)
↓
Structured Report