LatteReview

LatteReview 🤖☕

---

🚨 NEW in v1.4.0: Decision reviewers now also run on Perplexity’s pplx-decider and OpenAI’s gpt-6-luna through their new Decisions APIs, next to TypeSafe’s Jev. In our evaluation on 11,793 articles, pplx-decider ranked articles best, for about $0.05 per 1,000 abstracts. Perplexity’s Sonar works as an LLM reviewer too. See What’s New and Decision models vs LLM reviewers.

---

LatteReview is a powerful Python package designed to automate academic literature review processes through AI-powered agents. Just like enjoying a cup of latte ☕, reviewing numerous research articles should be a pleasant, efficient experience that doesn’t consume your entire day!

🆕 What’s New in v1.4.0

• Perplexity’s Decisions API: decision reviewers run on Perplexity’s pplx-decider with SystemOneProvider(backend="perplexity") (PERPLEXITY_API_KEY), or through OpenRouter with SystemOneProvider(backend="openrouter", model="perplexity/pplx-decider-v1.1-27b").
• OpenAI’s Decisions API: SystemOneProvider(backend="openai") runs gpt-6-luna through OpenAI’s new Decisions API. The provider translates OpenAI’s request format, so every decision reviewer works unchanged. Questions that OpenAI declines to answer come back as None instead of stopping the run.
• Three decision models compared on 11,793 articles: pplx-decider ranked best (mean AUC 0.895, best on 7 of 9 datasets, about $0.05 per 1,000 abstracts), then Jev (0.878, $0.06) and the v1 LLM reviewers (0.828). gpt-6-luna’s Decisions API scored 0.779 ($0.16): on long, multi-part criteria it returned 0.00 for almost every article. See the comparison.
• Perplexity’s models as LLM reviewers: LiteLLMProvider(model="perplexity/sonar") reaches Sonar, and other vendors’ models such as perplexity/openai/gpt-6-luna, through Perplexity’s new Agent API (Perplexity no longer serves Sonar as chat completions). About $0.20 per 1,000 abstracts with Sonar, without web search.
• Correct costs for per-request fees: LiteLLMProvider now uses the cost the API reports (OpenRouter and Perplexity), so fees such as Sonar’s search fee are counted. Before, a Sonar request through OpenRouter was recorded at about 1/18 of its price.
• Same input, same answer: decision reviewers leave the workflow’s Review Task ID line out of the model’s input. With it, a borderline article’s answer could depend on its row number.

A decision-model + LLM pipeline in a few lines: Perplexity’s pplx-decider screens every article, and only the articles it is unsure about go to an LLM (set PERPLEXITY_API_KEY and OPENAI_API_KEY):

import asyncio
import pandas as pd
from lattereview.providers import SystemOneProvider, OpenAIProvider
from lattereview.agents import DecisionTitleAbstractReviewer, TitleAbstractReviewer
from lattereview.workflows import ReviewWorkflow

inclusion = {1: "The study must involve CT scans.", 2: "The study must use deep learning."}
exclusion = {1: "The study must not include PET scans."}

decider = DecisionTitleAbstractReviewer(
    provider=SystemOneProvider(backend="perplexity"),  # or SystemOneProvider() for TypeSafe's Jev, backend="openai"
    name="Decider", inclusion_criteria=inclusion, exclusion_criteria=exclusion,
)
llm = TitleAbstractReviewer(
    provider=OpenAIProvider(model="gpt-6-luna"),
    name="LLM", inclusion_criteria=str(inclusion), exclusion_criteria=str(exclusion),
)

workflow = ReviewWorkflow(workflow_schema=[
    {"round": "A", "reviewers": [decider], "text_inputs": ["title", "abstract"]},
    {"round": "B", "reviewers": [llm], "text_inputs": ["title", "abstract"],
     "filter": lambda row: pd.isna(p := row["round-A_Decider_include_probability"]) or 0.1 <= p < 0.9},  # uncertain or unanswered
])
results = asyncio.run(workflow(pd.read_csv("articles.csv")))  # columns: title, abstract

Try it in the notebooks: Screening with Jev, pplx-decider and gpt-6-luna · Hybrid Jev + LLM review with measurements · Decision Models docs.

Existing reviewers and defaults work as before. See the CHANGELOG for the full list.

What Was New in v1.3.0

• Decision reviewers with Jev: LatteReview can now review with System One decision models such as TypeSafe’s Jev, which answer typed questions with probabilities instead of generating text. DecisionTitleAbstractReviewer, DecisionScoringReviewer and the generic DecisionReviewer work in any ReviewWorkflow, next to LLM reviewers.
• One provider, any backend: SystemOneProvider works with TypeSafe and OpenRouter, and with any other /v1/systemone server via base_url, including a self-hosted OpenJev model on your own machine.
• Evaluated at full scale: on all 11,793 articles of LatteReview’s evaluation datasets, Jev ranked articles better than the v1 LLM reviewers on every dataset (mean AUC 0.88 vs 0.83), for about $0.06 per 1,000 articles. See the evaluation, including where Jev falls short.
• Thresholds for a target recall: suggest_threshold fits a probability cutoff on labeled data.
• Hybrid workflows: let Jev screen everything and send only uncertain articles to an LLM (tutorial).

What Was New in v1.2.0

• Current models: tested with OpenAI GPT-6 (gpt-6-astra, gpt-6-sol, gpt-6-luna) and GPT-5.x, Anthropic Claude Opus 5.5, Sonnet 5, Haiku 4.5 and Fable 5.1, and Google Gemini 3.x (gemini-3.8-flash, gemini-3.5-flash-lite). Older models such as gpt-4o-mini and gemini-2.5-flash keep working.
• No more rejected-parameter errors: if a model rejects a setting in model_args (e.g., temperature on GPT-6 or Claude 5, or max_tokens on OpenAI reasoning models), LatteReview drops or renames it with a one-time warning and retries. If a reasoning model runs out of tokens before finishing its answer, the call is retried without the limit.
• New default models: OpenAIProvider and LiteLLMProvider default to gpt-6-luna, GoogleProvider to gemini-3.8-flash, and OllamaProvider to qwen3.8:27b. Pass model= to choose another.
• Better local models: OllamaProvider constrains answers to the reviewer’s JSON schema, maps reasoning_effort to Ollama’s think setting, passes other model_args (e.g., top_p) as model options instead of failing, and close() works again. Tested with qwen3.8:27b on a 32 GB Apple Silicon Mac.
• More accurate costs: computed from the token usage each API reports, including hidden reasoning tokens.
• Python 3.10 or later is now required. On Python 3.9, pip installs 1.1.1.

🎯 Key Features

• Multi-agent review system with customizable roles and expertise levels for each reviewer
• Support for multiple review rounds with hierarchical decision-making workflows
• Review diverse content types including article titles, abstracts, custom texts, and even images using LLM-powered reviewer agents
• Define reviewer agents with specialized backgrounds and distinct evaluation capabilities (e.g., scoring or concept abstraction or custom reviewers of your own preferance)
• Create flexible review workflows where multiple agents operate in parallel or sequential arrangements
• Enable reviewer agents to analyze peer feedback, cast votes, and propose corrections to other reviewers’ assessments
• Enhance reviews with item-specific context integration, supporting use cases like Retrieval Augmented Generation (RAG)
• Broad compatibility with LLM providers through LiteLLM, including OpenAI and Ollama
• Model-agnostic integration supporting OpenAI, Gemini, Claude, Groq, DeepSeek, Perplexity, OpenRouter, and local models via Ollama
• NEW: Decision-model reviewers (Perplexity’s pplx-decider, TypeSafe’s Jev, OpenAI’s gpt-6-luna, or a self-hosted OpenJev) that return probabilities for fast, cheap screening
• High-performance asynchronous processing for efficient batch reviews
• Standardized output format featuring detailed scoring metrics and reasoning transparency
• Robust cost tracking and memory management systems
• Extensible architecture supporting custom review workflow implementation
• NEW: Support for RIS (Research Information Systems) file format for academic literature review

💾Installation

pip install lattereview

LatteReview requires Python 3.10 or later. Please refer to the installation guide for detailed instructions.

🚀 Quick Start and Documentation

LatteReview enables you to create custom literature review workflows with multiple AI reviewers. Each reviewer can use different models and providers based on your needs. Below is a working example of how you can use LatteReview for doing a quick title/abstract review with two junior and one senior reviewers (all AI agents)! And this is just the beginning! Beyond study screening, LatteReview can handle data abstraction, customized pipelines, image analysis, and much more. Explore the Tutorials for more examples!

Please refer to the Quick Start page and Documentation page for detailed instructions.

The example below is fully self-contained: set your API keys, install the package, and run it as-is. It uses one OpenAI and one Gemini model, so it needs OPENAI_API_KEY and GEMINI_API_KEY (in a .env file or exported in your shell). You can swap in any LiteLLM-supported model — see Model Support.

from lattereview.providers import LiteLLMProvider
from lattereview.agents import TitleAbstractReviewer
from lattereview.workflows import ReviewWorkflow
import pandas as pd
import asyncio
from dotenv import load_dotenv

# Load environment variables (e.g., OPENAI_API_KEY, GEMINI_API_KEY) from a .env file
load_dotenv()

# First Reviewer: Conservative approach
reviewer1 = TitleAbstractReviewer(
    provider=LiteLLMProvider(model="gpt-6-luna"),
    name="Alice",
    backstory="a radiologist with expertise in systematic reviews",
    inclusion_criteria="The study must focus on applications of artificial intelligence in radiology.",
    exclusion_criteria="Exclude studies that are not peer-reviewed or not written in English.",
    model_args={"reasoning_effort": "low"},
)

# Second Reviewer: More exploratory approach
reviewer2 = TitleAbstractReviewer(
    provider=LiteLLMProvider(model="gemini/gemini-3.8-flash"),
    name="Bob",
    backstory="a computer scientist specializing in medical AI",
    inclusion_criteria="The study must focus on applications of artificial intelligence in radiology.",
    exclusion_criteria="Exclude studies that are not peer-reviewed or not written in English.",
    model_args={"reasoning_effort": "low"},
)

# Expert Reviewer: Resolves disagreements
expert = TitleAbstractReviewer(
    provider=LiteLLMProvider(model="gpt-6-sol"),
    name="Carol",
    backstory="a professor of AI in medical imaging",
    inclusion_criteria="The study must focus on applications of artificial intelligence in radiology.",
    exclusion_criteria="Exclude studies that are not peer-reviewed or not written in English.",
    model_args={"reasoning_effort": "high"},
    additional_context="Alice and Bob disagree with each other on whether or not to include this article. You can find their reasonings above.",
)

# Define workflow
workflow = ReviewWorkflow(
    workflow_schema=[
        {
            "round": 'A',  # First round: Initial review by both reviewers
            "reviewers": [reviewer1, reviewer2],
            "text_inputs": ["title", "abstract"]
        },
        {
            "round": 'B',  # Second round: Expert reviews only disagreements
            "reviewers": [expert],
            "text_inputs": ["title", "abstract", "round-A_Alice_output", "round-A_Bob_output"],
            "filter": lambda row: row["round-A_Alice_evaluation"] != row["round-A_Bob_evaluation"]
        }
    ]
)

# Prepare your data: a DataFrame (or .csv/.xlsx/.ris file path) with 'title' and 'abstract' columns
data = pd.DataFrame(
    {
        "title": [
            "Deep learning for automated detection of pneumonia on chest radiographs",
            "Effects of mindfulness meditation on stress levels in college students",
        ],
        "abstract": [
            "We developed a convolutional neural network to detect pneumonia on chest X-rays.",
            "A randomized trial of mindfulness training in 200 undergraduates reduced stress.",
        ],
    }
)
# Or load from a file: data = pd.read_excel("articles.xlsx")

results = asyncio.run(workflow(data))  # Returns a pandas DataFrame with all original and output columns

# Save results
results.to_csv("review_results.csv", index=False)

🎯 Decision Models

Decision reviewers use a decision model instead of an LLM. Three models are supported, all through SystemOneProvider:

Model	Provider	API key	Mean AUC*	Cost per 1,000 abstracts*	Notes
Perplexity’s pplx-decider	SystemOneProvider(backend="perplexity")	PERPLEXITY_API_KEY	0.895	$0.05	Best in our evaluation; open weights (Apache-2.0); bills the abstract once per question
TypeSafe’s Jev	SystemOneProvider() (default)	TYPESAFE_API_KEY	0.878	$0.06	Nearly free to add questions; uncertain answers vary slightly between runs
OpenAI’s gpt-6-luna	SystemOneProvider(backend="openai")	OPENAI_API_KEY	0.779	$0.16	Weak on long, multi-part criteria; may decline single questions (answer = None)

* DecisionTitleAbstractReviewer on all 11,793 articles of LatteReview’s evaluation datasets; the v1 LLM reviewers scored 0.828. Jev and pplx-decider are also available through OpenRouter (backend="openrouter", OPENROUTER_API_KEY).

from lattereview.providers import SystemOneProvider
from lattereview.agents import DecisionTitleAbstractReviewer

decider = DecisionTitleAbstractReviewer(
    provider=SystemOneProvider(backend="perplexity"),  # pplx-decider-v1.1-27b
    name="Decider",
    inclusion_criteria={1: "The study must involve CT scans.", 2: "The study must use deep learning."},
    exclusion_criteria={1: "The study must not include PET scans."},
)
# Use it in a ReviewWorkflow like any reviewer. Columns: evaluation (1-5), include_probability,
# confidence, criteria (per-criterion probabilities) and reasoning (generated from the probabilities).

Decision models vs LLM reviewers. An LLM reviewer writes its answer and a reasoning; a decision model writes nothing and returns a probability for every allowed answer, so its answers are always on-schema and can be thresholded for a target recall. Decision models are fast (under a second per article) and cheap (about $0.05-0.16 per 1,000 abstracts), and in our evaluation pplx-decider and Jev ranked articles better than the v1 LLM reviewers. But they read text only, write no reasoning, do not do multi-step reasoning, and a 0.5 cutoff is often too strict for long, multi-part criteria: rank by include_probability or fit a cutoff per model with suggest_threshold. A good pattern is to let a decision model screen everything and send only uncertain articles to an LLM. Read Decision Models for how they work, backends, costs, self-hosting and the full evaluation.

🔌 Model Support

LatteReview offers flexible model integration through multiple providers:

• LiteLLMProvider (Recommended): Supports OpenAI, Anthropic (Claude), Gemini, Groq, DeepSeek, Perplexity (Sonar), OpenRouter, and more
• OpenAIProvider: Direct integration with OpenAI and Gemini APIs
• GoogleProvider: Direct integration with Gemini through Google’s google-genai SDK
• OllamaProvider: Optimized for local models via Ollama
• SystemOneProvider: decision models: TypeSafe’s Jev, Perplexity’s pplx-decider and OpenAI’s gpt-6-luna (Decisions APIs), via TypeSafe, Perplexity, OpenAI, OpenRouter, or any /v1/systemone server (e.g., a self-hosted OpenJev)

If you don’t pass a model, OpenAIProvider and LiteLLMProvider use gpt-6-luna, GoogleProvider uses gemini-3.8-flash, and OllamaProvider uses qwen3.8:27b (run ollama pull qwen3.8:27b first). For Claude, pass e.g. LiteLLMProvider(model="anthropic/claude-sonnet-5") or "anthropic/claude-haiku-4-5" for a cheaper option. For Perplexity, pass LiteLLMProvider(model="perplexity/sonar") (PERPLEXITY_API_KEY); LatteReview sends it through Perplexity’s Agent API.

Note: Models should support async operations and structured JSON outputs for optimal performance.

Model compatibility

Newer reasoning models reject some request parameters that older models accept. OpenAI’s GPT-5/GPT-6 families and o-series models reject max_tokens (they use max_completion_tokens) and non-default temperature/top_p. Anthropic’s Claude Opus 4.7+, Claude 5 and Fable models reject temperature, top_p and top_k. LatteReview handles this for you. If a model rejects a parameter in model_args, it is dropped (or max_tokens is sent as max_completion_tokens) with a one-time warning. If a reasoning model runs out of tokens before finishing its answer, the call is retried without the token limit. Models that accept these parameters are called exactly as before, so existing code keeps working.

For reasoning models, we recommend leaving out max_tokens and temperature and using reasoning_effort (e.g., "low" for screening, "high" for an expert reviewer) instead. See Model compatibility in the docs for details.

📖 Documentation

Full documentation and API reference are available at: https://pouriarouzrokh.github.io/LatteReview

🎓 Tutorials

✅ TitleAbstractReviewer: 🔸1. A simple task of abstract screening based on 1-5 scoring + inclusion and exclusion criteria ✅ AbstractionReviewer: 🔸1. Data abstraction from abstracts/manuscripts ✅ ScoringReviewer: 🔸1. A simple task of abstract screening based on custom scoring by multiple agents 🔸2. Question answering with RAG (Retrieval Augmented Generation) 🔸3. Image analysis by LatteReview
✅ Custom Reviewer: 🔸1. How to Customize the AbstractReviewer Agent for Your Needs 🔸2. Chat with the agents and other base functionalities 🔸3.: Combination of differnet agents for a comprehensive literature review ✅ Decision Reviewers (Jev, pplx-decider, gpt-6-luna): 🔸1. Screening, scoring and categorical extraction with Jev, and the same screening with pplx-decider and gpt-6-luna 🔸2. Hybrid review: Jev screens everything, an LLM handles the uncertain articles

🛣️ Roadmap for Future Features

• Implementing LiteLLM to add support for additional model providers
• Draft the package full documentation
• Enable agents to return a percentage of certainty
• Enable agents to be grounded in static references (text provided by the user)
• Enable agents to be grounded in dynamic references (i.e., recieve a function that outputs a text based on the input text. This function could, e.g., be a RAG function.)
• Support for image-based inputs and multimodal analysis
• Development of AbstractionReviewer class for automated paper summarization
• Showcase how AbstractionReviewer class could be used to analyse the literature around a certain topic.
• Adding a tutorial example and also a section to the docs on how to create custom reviewer agents.
• Adding a TitleAbstractReviewer agent and adding a tutorial for it.
• Evaluating LatteReview.
• Writing the white paper for the package and public launch
• Addign support for RIS files.
• Adding support for models without structured-output (json_schema) capability via an automatic JSON-mode fallback (e.g., DeepSeek).
• Supporting the newest reasoning models (GPT-6, Claude 5, Gemini 3.x) with automatic handling of parameters they reject.
• Supporting System One decision models (Jev) for fast, probability-based screening, including hybrid Jev + LLM workflows.
• Supporting Perplexity’s and OpenAI’s Decisions APIs, and Perplexity’s Sonar LLMs.
• Development of a no-code web application
• (for v>) Adding conformal prediction tool for calibrating agents on their certainty scores
• (for v>2.0.0) Adding a dialogue tool for enabling agents to seek external help (from helper agents or parallel reviewer agents) during review.
• (for v>2.0.0) Adding a memory component to the agents for saving their own insights or insightful feedback they receive from the helper agents.

👨‍💻 Author

Pouria Rouzrokh, MD, MPH, MHPE Medical Practitioner and Machine Learning Engineer Incoming Radiology Resident @Yale University Former Data Scientist @Mayo Clinic AI Lab

❤️ Support LatteReview

If you find LatteReview helpful in your research or work, consider supporting its continued development. Since we’re already sharing a virtual coffee break while reviewing papers, maybe you’d like to treat me to a real one? ☕ 😊

Ways to Support:

• Become my sponsor on GitHub
• Treat me to a cup of coffee on Ko-fi ☕
• Star the repository to help others discover the project
• Submit bug reports, feature requests, or contribute code
• Share your experience using LatteReview in your research

📜 License

LatteReview is licensed under the PolyForm Noncommercial License 1.0.0 from version 1.4.1 on. You may use, change and share it for any noncommercial purpose, including personal research and use by universities, hospitals, public research organizations, charities and government institutions. Commercial use needs a separate license: please contact the author to ask for one. Releases up to and including 1.4.0 remain under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

🤝 Contributing

I welcome contributions! Please feel free to submit a Pull Request.

Acknowledgement

I would like to express my heartfelt gratitude to Moein Shariatnia for his invaluable support and contributions to this project.

📚 Citation

If you use LatteReview in your research, please cite our paper:

@misc{rouzrokh2025lattereview,
    title={LatteReview: A Multi-Agent Framework for Systematic Review Automation Using Large Language Models},
    author={Pouria Rouzrokh and Moein Shariatnia},
    year={2025},
    eprint={2501.05468},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}