(8/n) Get started with LOTUS for your own LLM data processing tasks and build with us ❤️

🌟 LOTUS code: https://github.com/lotus-data/lotus 📊 Benchmark code: https://github.com/lotus-data/lotus/tree/main/benchmarks 📝 Full blog https://liana313.github.io/blog/lotusplan.html 💬 Discord: https://discord.gg/ZWQBurm5bt

(2/n) Over a year ago, LOTUS introduced semantic operators — declarative, AI-based primitives (sem_filter, sem_map, sem_agg, sem_join) for bulk-processing unstructured data with LLMs.

Think the performance and reliability of relational operators, but for messy text and LLM-powered transformations.

More on semantic operators here: https://arxiv.org/abs/2407.11418

🚀 Beyond excited to share we're releasing LOTUSPlan, a new API & optimizer for higher performance LLM-powered data processing, from our team at Berkeley & Stanford.

LOTUS now lets you write your LLM-based queries and optimize them for up to 2.4× lower cost and 4.6× higher accuracy for tasks like, agent trace analysis, LLM-judge evals, RAG, document extraction and deep research.

✨Checkout our our new blog: https://liana313.github.io/blog/lotusplan.html

🧵

(7/n) DeepResearch

And LOTUS powers real systems, like DeepScholar, our open DeepResearch pipeline, which synthesizes 100s of articles into well-cited reports — competitive with OpenAI's DeepResearch and ~2× faster!

More on DeepScholar here:

(3/n) LOTUSPlan represents the next evolution of the LOTUS API toward making LLM-based data processing easier, cheaper, and more accurate.

Using global planning and lazy execution for your LLM queries, LOTUSPlan significantly boosts performance. And the results are very exciting...

(4/n) Agent Trace Analysis:

🤖LOTUS allows you to analyzing agent traces to find key failure modes with a simple 2-operator LOTUS query. Optimized with LOTUSPlan, we achieve 4.6× higher accuracy and 1.4× lower cost. This makes it faster and easier for you to extract critical insights on your agents.

(6/n) RAG

📚LOTUS let's you implement RAG as a simple query and LOTUSPlan jointly optimizes the whole pipeline for 50% higher answer accuracy at no added cost.

(5/n) LLM-judge evals:

⚖️ When running LLM-judge evals at scale, LOTUSPlan jointly optimizes the judge prompt (with the power of GEPA!) and execution plan to achieve 9.5% higher accuracy and 2.5× lower cost. Cheaper, more reliable evals.

(9/n) Huge thanks to @harshitgupta412 for helping to lead this release and to amazing advisors @istoica05 @guestrin @matei_zaharia — and the whole LOTUS team and community. There's so much more to come!

@lianapatel_ Lotus 🐐

@lianapatel_ Yes🙌