Overview
A.DOT is a clear engineering prototype: it shows strong gains on HybridQA dev with concrete modules for validation, caching, and remediation, but it remains a research prototype pending large-scale enterprise tests and user studies.
Citations0
Evidence Strength0.78
Confidence0.80
Risk Signals10
Trust Signals
Findings with numeric evidence: 3/3
Findings with evidence refs: 3/3
Results with explicit delta: 4/4
Reproducibility
Status: Partial assets available
Open source: Partial
At A Glance
Cost impact: 60%
Production readiness: 70%
Novelty: 66%
Why It Matters For Business
A.DOT reduces over-retrieval and exposes verifiable evidence for each step, cutting unnecessary data exposure and improving multi-step QA accuracy—helpful for compliance-heavy enterprise queries.
Who Should Care
Summary TLDR
A.DOT compiles a user's natural-language question into a directed acyclic graph (DAG) of atomic subqueries that target either a relational database or a vector store. A single LLM pass generates the plan, which is validated, cached, and executed in parallel. The executor passes only needed IDs (not full payloads) between steps, records evidence trails, and invokes DataOps agents to diagnose and repair plan errors. On the HybridQA dev set, A.DOT improves answer correctness by 14.8% and completeness by 10.7% over a strong RAG baseline. The system is a research prototype under enterprise evaluation.
Problem Statement
Enterprises need natural-language question answering over mixed data lakes (relational tables + document vectors). Existing RAG or tool-calling pipelines over-retrieve, leak data, and fail at multi-hop queries that alternate between structured and unstructured sources. A practical system must plan precise cross-source queries, run them efficiently, and produce verifiable evidence.
Main Contribution
A.DOT: an agentic DAG planner that compiles NL queries into atomic, source-targeted subqueries and executes them according to DAG dependencies.
Schema-aware plan validation plus a DataOps loop (diagnose, fix, replan) to detect and repair plan or execution errors.
Key Findings
A.DOT improves answer correctness and completeness on HybridQA dev vs Standard RAG.
DataOps system is critical—removing it drops correctness sharply.
Results
| Metric | Value | Baseline | Delta | Split / Dataset | Evidence | Evidence Ref |
|---|---|---|---|---|---|---|
| Answer Correctness | A.DOT 71.0 | Standard RAG 56.2 | +14.8 | HybridQA dev | Table 1: A.DOT vs baselines | Table 1 |
| Answer Completeness | A.DOT 73.0 | Standard RAG 62.3 | +10.7 | HybridQA dev | Table 1: A.DOT vs baselines | Table 1 |
What To Try In 7 Days
Prototype a DAG plan for 5 common cross-source NL queries in your data lake.
Add a light schema check that rejects plans referencing missing columns.
Log variable bindings (IDs) instead of full payloads when passing results across steps to reduce data leakage and token use.
Agent Features
Memory
Planning
Tool Use
Frameworks
Is Agentic
Yes
Architectures
Collaboration
Optimization Features
Token Efficiency
System Optimization
Inference Optimization
Reproducibility
Data URLs
Risks & Boundaries
Limitations
Evaluation limited to the HybridQA dev split; production datasets not yet reported.
Prototype relies on LLaMA-3-70B, which is compute-intensive for real-time deployment.
When Not To Use
Simple single-source Q&A where single-pass RAG already suffices.
Very low-compute environments that cannot host 70B models.
Failure Modes
Intermediate subquery failures (NL-to-SQL or NL-to-Vector modules fail).
Incorrect data-source assignment by the planner.