Agentic Academic Paper Assistant

An agentic, multi-stage academic paper assistant built with LangGraph / LangChain that can search, read, summarize, verify, and cite research papers in a reliable, production-oriented way.

---

✨ Key Features

• Multi-agent architecture with clear role separation
• Graph-based orchestration (LangGraph) instead of fragile chains
• Academic paper search (e.g. arXiv)
• PDF download & parsing
• Structured paper summaries (problem, method, contribution, limitations)
• Citation generation (APA / BibTeX / IEEE-ready)
• Long-term memory using VectorDB (Chroma)
• Failure-aware fallbacks (no results, PDF unavailable, low confidence)
• FastAPI service interface for deployment

---

🧠 System Design

This system is a role-based multi-agent system with centralized orchestration.

Instead of letting agents freely chat, the system enforces a structured workflow:

User Query
   ↓
Router / Planner Agent
   ↓
Paper Retrieval Agent
   ↓
PDF Reader / Summarizer Agent
   ↓
Citation & Verification Agent
   ↓
Memory (RAG) Update
   ↓
Final Response

This design improves:

• reliability
• debuggability
• termination guarantees
• production readiness

---

🧩 Agents & Responsibilities

1. Router / Planner Agent

• Interprets user intent
• Expands vague queries into academic language
• Decides which agents should run
• Sets constraints (top_k, year range, domain)

2. Paper Retrieval Agent

• Searches academic sources (e.g. arXiv)
• Deduplicates results
• Returns structured metadata

3. PDF Reader / Summarizer Agent

• Downloads PDFs when available
• Extracts key sections
• Produces structured summaries

4. Citation & Verification Agent

• Generates citations (BibTeX / APA)
• Verifies claims against sources
• Flags low-confidence outputs

5. Memory Agent (RAG)

• Stores previously seen papers
• Enables long-term recall
• Avoids repeated searches

---

🧠 Memory Design

The system separates short-term state and long-term memory:

• Short-term state (in-session): the LangGraph AgentState plus a small short-term memory layer:
  • messages: recent user + assistant turns (bounded, e.g. last 20)
  • last_papers, last_citations, last_intent, last_slots: last turn’s context, used for follow-ups like “cite those in APA” or “download the second one”
  • Implemented in src/agent/state.py and src/agent/short_term_memory.py
• Long-term memory (persistent): VectorDB (Chroma)
  • One paper = one vector entry
  • Metadata stored alongside embeddings

This enables:

• conversational follow-ups without repeating the full query
• persistent knowledge across sessions
• faster responses for repeated topics
• incremental system learning

---

🚑 Failure Handling & Fallbacks

The agent is designed to fail gracefully:

• No search results → query expansion & retry
• PDF unavailable → abstract-only summarization
• Low citation confidence → uncertainty explicitly reported
• Tool failure → controlled retry or safe exit

This is a core 2025 agent design principle.

---

🛠️ Tech Stack

• Python
• LangGraph (graph-based agent orchestration)
• LangChain (LLM & tool abstractions)
• Chroma (vector database for memory)
• FastAPI (service layer)
• arXiv API (paper search)

---

🚀 Running the Project

# install dependencies
pip install -r requirements.txt

# run API server (stateless, one-shot requests)
uvicorn app:app --reload

CLI (multi-turn with short-term memory)

python main.py

In the CLI:

• Each message runs the LangGraph workflow.
• Short-term memory is carried across turns, so you can:
  • First: “Find recent papers about diffusion models.”
  • Then: “Cite those in APA.” (reuses last_papers instead of searching again)

Debug mode (print full agent state)

python main.py --debug

With --debug, the CLI prints the full final graph state (including intent, actions, papers, citations, memory logs) after each turn, which is useful when debugging routing or workflows.

---

📌 Design Philosophy

This project intentionally avoids:

• free-form agent chat
• uncontrolled agent loops
• "emergent" coordination without rules

Instead, it follows explicit workflows, role boundaries, and verification steps — aligning with current best practices in multi-agent research (2025).

---

📖 Motivation

Single LLM calls are powerful but unreliable for academic work.

This project explores how agentic workflows and multi-agent coordination can make LLM systems:

• more trustworthy
• more controllable
• more suitable for real research use

---

🔮 To do list

• Multi-source retrieval (Semantic Scholar, OpenReview)
• Improve the performance of search agent
• Summary Node(Read the Node According to the pdf)
• pdf upload
• Service Env set up
• UI Design

---

📄 License

MIT License

---

If you are interested in agentic systems, academic assistants, or production-grade LLM workflows, feel free to explore or contribute.