BatchDAG: LLM-планировщик строит граф операций и режет вызовы модели в 47 раз

In July 2026, researchers published a system called BatchDAG on arXiv, in which a language model builds an execution graph for a query against enterprise data, while a separate engine executes it; the approach cuts the number of model calls by up to 47x, and in production at Brevian.ai it processes queries across 50000+ meetings in under 60 seconds.

How BatchDAG Works

BatchDAG operates in two steps: first, a large language model (LLM) generates a typed directed acyclic graph (DAG) of operations, then a deterministic engine executes it with parallelism across topological waves. The graph's nodes are SQL queries, semantic search, in-memory transformations, parallel branching (fan-out), and single-item analyses, with data passed between them as structured JSON.

BatchDAG's planner produces a correct graph in 98,8% of cases — measured across 300 planning calls, according to the arXiv paper.

• Reduction in LLM calls — up to 47 times thanks to entity-aware batching
• Correct DAG — in 98,8% of 300 planner calls
• Scale — 50000+ meetings processed in under 60 seconds
• Query cost — $0,02–$0,24 at published GPT-5.1 prices
• Answer quality — 3,74/5 vs. 3,25/5 for the expert pipeline

Why Fewer Hallucinations

Structured JSON intermediate results reduce hallucinations by 27% compared to text-based summaries — this was shown in the authors' controlled ablation experiment (paired t-test, p=0.107, n=12). Intermediate data with an explicit schema doesn't let the model "make up" facts between steps.

The key optimization is entity-aware batching: before parallel branching, the system groups rows by logical entity, which is what delivers the up to 47x reduction in LLM calls. This also improves verifiability: the share of answers backed by a citation from the transcript reached 77% for BatchDAG, versus 46–60% for baseline approaches.

How It Performed in Testing

In controlled experiments across 12 meeting-transcript queries, BatchDAG scored 3,74/5, beating the expert pipeline (3,25/5) and a ReAct agent (3,09/5, p<0.01). The authors emphasize that the system's goal isn't accuracy for its own sake, but universality.

"BatchDAG is less about an accuracy gain over manually optimized

pipelines, and more about being a universal orchestration layer that replaces a multitude of handwritten scripts with a single system," the arXiv paper states.

In production at Brevian.ai, the system runs on top of GPT-5.1 and keeps the cost per query in the $0,02–$0,24 range.

What This Means

BatchDAG shows a practical path for analytics over large volumes of enterprise data: instead of dozens of handwritten pipelines, a single system that translates a natural-language query into an execution plan on its own. For teams, this means faster, verifiable answers at a scale of 50000+ documents with response times under a minute — without manual tuning for every type of question.

Frequently Asked Questions

What Is BatchDAG?

BatchDAG is an orchestration layer in which an LLM builds a typed graph of operations (SQL, semantic search, transformations), while a deterministic engine executes it in parallel. The system was introduced in a paper on arXiv in July 2026.

How Much Does One Query Cost?

According to the authors, the cost of a single query ranges from $0,02 to $0,24 at published GPT-5.1 prices — while processing a corpus of 50000+ meetings in under 60 seconds.

How Does BatchDAG Differ from a ReAct Agent?

BatchDAG builds the full execution graph upfront and runs independent branches in parallel, whereas a ReAct agent acts step by step. In testing, this produced 3,74/5 versus 3,09/5 for ReAct and eliminates the linear delay from sequential tool calls.