LOTUS: Fast, Easy and Accurate LLM-Powered Data Processing

LOTUS is the framework that allows you to easily process your datasets, including unstructured and structured data, with LLMs. It provides an intuitive Pandas-like API, offers algorithms for optimizing your programs for up to 1000x speedups, and makes LLM-based data processing robust with accuracy guarantees with respect to high-quality reference algorithms.

LOTUS stands for LLMs Over Text, Unstructured and Structured Data, and it introduces semantic operators. Semantic operators extend the core philosophy of relational operators—designed for declarative and robust structured-data processing—to unstructured-data processing with AI. Semantic operators are expressive, allowing you to easily capture all of your data-intensive AI programs, from simple RAG, to document extraction, image classification, LLM-judge evals, unstructured data analysis, complex research-based synthesis and more.

For trouble-shooting or feature requests, please raise an issue and we'll get to it promptly. To share feedback and applications you're working on, you can send us a message on our community discord, or send an email (lianapat@stanford.edu).

Installation

Using uv (Recommended)

For the latest stable release:

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a new project or navigate to your existing project
uv add lotus-ai

For the latest features:

uv add git+https://github.com/lotus-data/lotus.git@main

Using pip

For the latest stable release:

conda create -n lotus python=3.10 -y
conda activate lotus
pip install lotus-ai

For the latest features, you can alternatively install as follows:

conda create -n lotus python=3.10 -y
conda activate lotus
pip install git+https://github.com/lotus-data/lotus.git@main

Running on Mac

If you are running on mac and using pip, please install Faiss via conda:

CPU-only version

conda install -c pytorch faiss-cpu=1.8.0

GPU(+CPU) version

conda install -c pytorch -c nvidia faiss-gpu=1.8.0

If you're using uv, the faiss-cpu dependency will be handled automatically.

For more details, see Installing FAISS via Conda.

Quickstart

If you're already familiar with Pandas, getting started will be a breeze! Below we provide a simple example using sem_filter. Like all semantic operators, it is specified by a langex (natural language expression) parameterized by one or more column names in brackets — here {title}. The langex for a sem_filter is a NL predicate, which is a natural language expression that can be evaluated to a True/False value.

import pandas as pd
import lotus
from lotus.models import LM

# Configure the LM — export your API key before running (e.g. OPENAI_API_KEY)
lm = LM(model="gpt-4.1-nano")
lotus.settings.configure(lm=lm)

# A sample of GitHub-style issue titles from an open source project
issues = pd.DataFrame({
    "title": [
        "Fix typo in README",
        "Add dark mode support to dashboard",
        "Refactor entire auth system to use OAuth2",
        "Update copyright year in LICENSE",
        "Implement distributed transaction support across microservices",
        "Change button color on settings page",
        "Migrate database from Postgres 13 to 16 with zero downtime",
        "Add missing comma in error message",
        "Build custom query planner to replace third-party dependency",
        "Bump lodash to fix known CVE",
        "Support multi-region active-active replication",
        "Remove unused import in utils.py",
    ]
})

# Use sem_filter to find issues approachable for a first-time contributor
good_first_issues = issues.sem_filter(
    "The {title} describes a small, self-contained task that a new open source contributor could tackle without deep knowledge of the codebase"
)

print("Good first issues for new contributors:\n")
print(good_first_issues.to_string(index=False))

# Uncomment to print the total LM usage
# lm.print_total_usage()

Tutorials

Below are some short tutorials in Google Colab, to help you get started. We recommend starting with [1] Introduction to Semantic Operators and LOTUS, which will provide a broad overview of useful functionality to help you get started.

| Tutorial | Difficulty | Colab Link | |----------------------------------------------------|-----------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 1. Introduction to Semantic Operators and LOTUS | | | | 2. Failure Analysis Over Agent Traces | | | | 3. System Prompt Analysis with LOTUS | | | | 4. Processing Multimodal Datasets | | |

</div>

Key Concept: The Semantic Operator Model

LOTUS introduces the semantic operator programming model. Semantic operators are declarative transformations over one or more datasets, parameterized by a natural language expression, that can be implemented by a variety of AI-based algorithms. Semantic operators seamlessly extend the relational model, operating over tables that may contain traditional structured data as well as unstructured fields, such as free-form text. These modular language-based operators allow you to write AI-based pipelines with high-level logic, leaving optimizations to the query engine. Each operator can be implemented and optimized in multiple ways, opening a rich space for execution plans, similar to relational operators. To learn more about the semantic operator model, read the full research paper.

LOTUS offers a number of semantic operators in a Pandas-like API, some of which are described below. To learn more about semantic operators provided in LOTUS, check out the full documentation, run the colab tutorial, or you can also refer to these examples.

Semantic operators provide a unified API for both LLM-based primitives... | Operator | Description | |------------|-------------------------------------------------| | sem_map | Map each record using a natural language projection| | sem_filter | Keep records that match the natural language predicate |
| sem_extract | Extract one or more attributes from each row | | sem_agg | Aggregate across all records (e.g. for summarization) | | sem_topk | Order the records by some natural langauge sorting criteria | | sem_join | Join two datasets based on a natural language predicate |

... and embedding-based primitives: | Operator | Description | |------------|-------------------------------------------------| | sem_sim_join | Join two DataFrames based on semantic similarity | | sem_search | Perform semantic search the over a text column |

Supported Models

There are 3 main model classes

Lotus

Install / Use

README