Agentic DAG-Orchestrated Planner Framework for Multi-Modal, Multi-Hop Question Answering in Hybrid Data Lakes
A.DOT is an agentic planning framework for question answering over enterprise hybrid data lakes. It compiles natural language questions into executable DAG plans to support multi-hop reasoning across structured tables…
Summary
A.DOT is an agentic planning framework for question answering over enterprise hybrid data lakes. It compiles natural language questions into executable DAG plans to support multi-hop reasoning across structured tables and unstructured documents. It aims to improve correctness, completeness, latency, and auditability at the same time, while providing explicit evidence chains and data lineage.
Problem
- Existing enterprise RAG/tool-calling solutions usually perform brute-force retrieval separately over SQL databases and vector stores, then stitch results together afterward. This is inefficient, prone to over-retrieval and data leakage, and also lacks explicit multi-hop reasoning capability.
- Questions in hybrid data lakes often require repeated switching between tables and documents; without planned execution, models are prone to hallucination, choosing the wrong data source, and making answer provenance hard to trace.
- This matters because enterprise settings require not only correct answers, but also low latency, verifiability, auditability, and traceable data lineage.
Approach
- The core mechanism is to decompose a natural language question into multiple “atomic sub-questions” in a single LLM planning step, generating a dependency-aware DAG; each node targets only one data source type (SQL or vector store).
- The system first performs structural validation + semantic validation: checking schema validity, variable dependencies, acyclicity, whether intent is preserved, and whether aggregations/joins are executable; when issues arise, they are handed to DataOps for diagnosis, repair, or re-planning.
- During execution, the system runs according to DAG topological order, with independent nodes executed in parallel, and uses variable binding to pass only the minimal necessary intermediate results (such as document_id), reducing load and leakage risk.
- The framework also adds paraphrase-aware plan caching, enabling reuse of DAG plans for equivalent queries; at the same time, it records each step’s operations, inputs and outputs, and evidence sources to form a verifiable lineage/evidence trail.
Results
- On the HybridQA dev set (3,466 QA pairs), the main reported results for A.DOT are: Answer Correctness 71.0 and Answer Completeness 73.0.
- Compared with the strongest baseline, Standard RAG (Correctness 56.2, Completeness 62.3), A.DOT improves correctness by 14.8 percentage points and completeness by 10.7 percentage points.
- Other baseline results are: ReAct 40.2 / 44.3 and LLM Compiler 27.8 / 30.8, indicating that sequential tool calling alone or weaker DAG orchestration is insufficient for cross-modal multi-hop reasoning.
- In a 500-sample ablation study, full A.DOT reaches 71.8 / 74.3; removing DataOps drops performance to 60.0 / 61.8; removing Plan Validator yields 68.0 / 69.6; removing both gives 67.9 / 69.6.
- The paper also claims the system is being evaluated for integration with IBM Watsonx.data Premium, but no public quantitative results are yet provided for deployment performance, user studies, or large-scale performance testing.
Link
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.