brycewang-stanford/latte-review-guide

LatteReview Guide

Overview

LatteReview is a low-code Python package that uses AI agents to automate systematic literature reviews. It handles title/abstract screening, full-text assessment, data extraction, and PRISMA-compliant reporting — tasks that typically consume hundreds of researcher-hours. Supports multiple LLM backends (Anthropic, OpenAI, local models).

Installation

pip install lattereview

Core Workflow

Step 1: Initialize Review

from lattereview import ReviewProject

# Create a new review project
project = ReviewProject(
    name="ML in Medical Imaging Review",
    research_question="What deep learning architectures are used for "
                      "medical image segmentation?",
    inclusion_criteria=[
        "Uses deep learning for medical image segmentation",
        "Published in peer-reviewed venue",
        "Reports quantitative evaluation metrics",
    ],
    exclusion_criteria=[
        "Review/survey articles",
        "Non-English publications",
        "Conference abstracts only",
    ],
)

Step 2: Import Papers

# Import from various sources
project.import_papers("scopus_export.csv", source="scopus")
project.import_papers("pubmed_export.csv", source="pubmed")

# Or from a DataFrame
import pandas as pd
df = pd.read_csv("papers.csv")
project.import_from_dataframe(df,
    title_col="title",
    abstract_col="abstract",
    year_col="year",
)

print(f"Imported {project.total_papers} papers")

Step 3: AI Screening

from lattereview.agents import ScreeningAgent

# Configure screening agent
screener = ScreeningAgent(
    llm_provider="anthropic",
    model="claude-sonnet-4-20250514",
    criteria=project.inclusion_criteria,
    exclusion=project.exclusion_criteria,
)

# Title/abstract screening
results = screener.screen(
    project.papers,
    mode="title_abstract",
    confidence_threshold=0.7,
)

# Results include: decision, confidence, reasoning
for paper in results[:3]:
    print(f"{paper.title}")
    print(f"  Decision: {paper.decision} "
          f"(confidence: {paper.confidence:.2f})")
    print(f"  Reason: {paper.reasoning}")

Step 4: Data Extraction

from lattereview.agents import ExtractionAgent

extractor = ExtractionAgent(
    llm_provider="anthropic",
    fields={
        "architecture": "Deep learning architecture used",
        "dataset": "Medical imaging dataset",
        "modality": "Imaging modality (CT, MRI, X-ray, etc.)",
        "dice_score": "Best Dice similarity coefficient reported",
        "sample_size": "Number of images/patients",
    },
)

extracted = extractor.extract(project.included_papers)

# Export structured data
extracted.to_csv("extracted_data.csv")

Step 5: Generate Report

# PRISMA flow diagram
project.generate_prisma_diagram("prisma.png")

# Summary statistics
summary = project.summarize()
print(f"Screened: {summary['screened']}")
print(f"Included: {summary['included']}")
print(f"Excluded: {summary['excluded']}")

Configuration

# Use different LLM providers
screener = ScreeningAgent(
    llm_provider="openai",
    model="gpt-4o",
)

# Local models via Ollama
screener = ScreeningAgent(
    llm_provider="ollama",
    model="llama3",
    base_url="http://localhost:11434",
)

Dual-Reviewer Mode

# Simulate dual-reviewer screening for reliability
results = screener.dual_screen(
    project.papers,
    models=["claude-sonnet-4-20250514", "gpt-4o"],
    agreement_threshold=0.8,
)

# Papers with disagreement flagged for human review
conflicts = [p for p in results if p.agreement < 0.8]
print(f"{len(conflicts)} papers need human adjudication")

Use Cases

1. Systematic reviews: PRISMA-compliant literature reviews
2. Scoping reviews: Rapid evidence mapping
3. Meta-analysis preparation: Structured data extraction
4. Grant applications: Quick literature landscape assessment

References

• LatteReview GitHub
• LatteReview Documentation
• Rouzrokh, P. et al. (2024). "LatteReview: AI-Assisted Systematic Literature Reviews."