Beyond Quacking: Deep Integration of Language Models and RAG into DuckDB (www.vldb.org)

Researchers announced FlockMTL, an extension that embeds large language model (LLM) capabilities and retrieval-augmented generation (RAG) directly into a DBMS (demonstrated for DuckDB). The goal is to simplify knowledge-intensive analytics that must combine structured tables and unstructured text by avoiding complex external orchestration, data movement, and manual LLM context management. By making LLMs a first-class part of the database stack, developers can express RAG pipelines in plain SQL and keep retrieval, fusion, and reasoning co-located with the data. Technically, FlockMTL exposes LLM-powered scalar and aggregate functions so you can run per-row and grouped LLM inference and even chain predictions across tuples. It adds data-fusion primitives for hybrid search (structured + unstructured) and implements database-style optimizations—batching and meta-prompting—to improve throughput and context efficiency. Crucially, it introduces PROMPT and MODEL as SQL DDL objects alongside TABLE, enabling resource-independence and planner-friendly management of prompts and model endpoints. The result is a more declarative, optimizable path for building RAG applications inside the DBMS, reducing engineering overhead and latency while improving correctness and scalability for production LLM-infused analytics.