Agentic Research Paper Analyzer

A production-style workflow that turns research PDFs into structured, evidence-backed outputs (summary, findings, results, limitations, replication notes, critique) using LangGraph for orchestration and LangChain for structured output + tool calling. Runs locally with Ollama and persists a paper-specific retrieval index for Q&A.

Why this exists (real-world problem)

Teams waste hours on repetitive “paper triage” work: extracting core claims, checking evidence, capturing experimental setup, and writing review notes. This project automates that workflow with repeatable structure and traceability, so the output is usable for: literature reviews, reproducibility checks, technical due diligence, and internal knowledge capture.

What it delivers

• Goal-driven analysis: You provide an “analysis goal” (e.g., replication checklist + critique) and the system plans tasks accordingly.
• Structured outputs: Each task returns a validated schema (Pydantic) instead of unstructured text, reducing post-processing and brittle parsing.
• Evidence grounding via retrieval (RAG): Tasks are executed using retrieved passages from the paper, enabling citations/traceability in outputs.
• Local model execution: Uses ChatOllama for local inference and supports tool calling + structured output capabilities.

Architecture (How it Works)

This system is implemented as a stateful agent workflow using LangGraph.

The pipeline separates planning, execution, and validation into explicit stages:

1. Loader — parses the PDF and extracts raw text and metadata
2. Planner (Analyzer) — uses structured output to decompose the user’s goal into concrete analysis tasks
3. Workers (fan‑out) — execute tasks in parallel using retrieval‑augmented generation (RAG)
4. Reviewer — validates that extracted claims are grounded in evidence
5. Compiler — aggregates validated outputs into a final structured report

LangGraph’s fan‑out / fan‑in execution model enables parallel task execution with deterministic state merging, which closely mirrors how research analysis is performed.

Pipeline

1. Loader: reads PDF → text + metadata
2. Planner (Analyzer): generates a task plan using structured output
3. Workers (fan-out): execute each task in parallel, retrieving evidence from the indexed paper (fan-out/fan-in)
4. Reviewer: checks grounding and flags issues
5. Compiler: assembles a final JSON report

Tech stack

• LangGraph (StateGraph) for stateful orchestration and fan-out/fan-in patterns
• LangChain for structured outputs and retrieval integration
• Ollama via langchain-ollama (ChatOllama) for local LLM inference
• Chroma for persisted vector index per paper
• FastAPI backend + Streamlit UI

Quickstart

# 1) Start Ollama and pull models
ollama serve
ollama pull llama3.2
ollama pull nomic-embed-text

# 2) Install deps
pip install -r requirements.txt

# 3) Run backend
uvicorn app.main:app --reload --port 8000

# 4) Run UI
streamlit run app/streamlit_app.py