Unified Query and Workflow Model

• Unified Query and Workflow Model is an integrated framework that merges information-seeking queries with automated execution workflows, enhancing scalability and robustness.
• It employs formal abstractions to represent queries as first-class objects within workflows, enabling seamless cross-stage optimizations and adaptive operator synthesis.
• The model leverages diverse formalisms such as CQL, DAG, and Petri-net to achieve dynamic execution, improved performance, and interoperability across heterogeneous modalities.

A unified query and workflow model is an architectural and algorithmic framework that treats information-seeking (querying) and task execution (workflow) as elements of a single, formally integrated system, rather than as isolated components or serial stages. Such models have emerged in multiple research domains—data-centric AI, semantic web, business process automation, foundation models for biosignal intelligence, and dialogue systems—to address scalability, robustness, abstraction, and seamless automation across heterogeneous modalities, data types, services, and user intents.

1. Formal Abstractions Underlying Unified Query and Workflow Models

Unified query and workflow models adopt an abstraction in which user intent (expressed as a query) and the sequence of processing steps (workflow) belong to a shared algebraic or procedural representation. This paradigm formalizes queries as first-class objects within workflow descriptions and enables propagation of context, constraints, and optimizations across data acquisition, transformation, orchestration, service invocation, and result presentation (Li et al., 2 Feb 2026, Kohn et al., 2023, Awad et al., 24 Jan 2025, Ojha, 2012, Liang et al., 18 Dec 2025).

Tabular Overview:

Formalism	Query Representation	Workflow Representation
CQL-based (BEST)	Event triggers/streams	CQL statements as rules
Operator DAG (DataFlow)	Operator input/output schemas	DAG of operators (pipelines)
SQL/Analytics (DashQL)	SQL clause, visualization	Script of sequenced operators
Petri-Net (Semantic Web)	RDF triples + constraints	Petri net transitions
Prompt-driven (Dialogue, LLM)	NL query decomposed by prompt	Generation/planning via prompting

In unified models, the query is not only a request for information but also a specification for workflow construction, operator selection, or service invocation. The boundary between data discovery, preparation, orchestration, and delivery is thus erased, supporting more abstract and reconfigurable automation pipelines (Liang et al., 18 Dec 2025, Li et al., 2 Feb 2026).

2. System Architectures and Key Components

Unified query and workflow systems are instantiated via composite architectures in which each component operates over shared intermediate representations, often adopting modular and extensible pipeline or event-driven designs:

• Operator graph (DataFlow): Operators $O$ are functions over table-like datasets; pipelines are DAGs, and the DataFlow-Agent can synthesize pipeline graphs automatically from user intent via operator search, sequencing, and verification (Liang et al., 18 Dec 2025).
• Stream-table duality (BEST): All business process events are modeled as raw event streams and updatable state tables; CQL queries translate both procedural and declarative notations (e.g., BPMN, DCR) into a common execution substrate (Awad et al., 24 Jan 2025).
• Scripted workflow grammars (DashQL): Unified SQL+workflow grammars that add workflow and visualization primitives (FETCH, LOAD, VISUALIZE) to SQL, enabling end-to-end task graphs and cross-stage optimizations (Kohn et al., 2023).
• Prompt and plan generative (WorkflowLLM): LLMs (e.g., WorkflowLlama) are fine-tuned to emit both a high-level task plan and commented code given a user query and API documentation, learning to map directly from intent to branched executable workflows (Fan et al., 2024).
• Unified multimodal query sets (PanLUNA): In biosignal intelligence, modality-specific channels (e.g., EEG, ECG, PPG) are embedded as a unified query set, with sensor-type embeddings to guide early fusion and robustness to missing modalities (Zelic et al., 5 Apr 2026).
• Dialogue and information retrieval (UniRQR): A single parameter-sharing model conducts retrieval decision, query generation, and response formulation in a multitask learning loop using prompts to select the task mode (Hu et al., 2024).

3. Workflow Construction, Execution Semantics, and Optimization

The process by which unified models translate queries into workflows typically involves several distinct but coordinated phases that depend on the abstraction layer:

• Semantic Decomposition and Operator Synthesis: Natural-language queries are decomposed into sub-queries or processing steps (e.g., via LLM-based classifiers, chain-of-thought reasoning) (Liang et al., 18 Dec 2025, Fan et al., 2024, Li et al., 2 Feb 2026). Each sub-query is mapped to atomic operators or service calls, and a graph or plan is constructed, either as a DAG of operators (DataFlow) or as Python-style code (WorkflowLlama).
• Event-Driven Execution (BEST): Events (task status changes, service invocations) are processed by event/stream handlers, with CQL expressing transformations over both the event streams and state tables. Process modeling languages (BPMN, DCR) are compiled into CQL rules over the same stateful event model, allowing hybrid paradigms to interoperate (Awad et al., 24 Jan 2025).
• Formal Correctness and Guarantees: Models are frequently cast in Petri-net or algebraic forms, guaranteeing properties such as completeness of identification (if a satisfying service chain exists, it is discovered), and optimality of composition (highest utility chain under negotiation/auction) (Ojha, 2012).
• Cross-Stage Optimization: In SQL+workflow languages (DashQL), visualization contexts inform upstream query predicates, enabling predicate pushdown, materialization, and adaptive aggregation (M4/AM4). Holistic optimizations are possible due to the integrated AST and metadata propagation (Kohn et al., 2023).

4. Modality Unification and Robustness Mechanisms

Unified query-workflow models facilitate seamless handling of heterogeneous modalities, missing channels, incomplete information, and operator selection:

• Unified Key/Query Pools (PanLUNA): Each sensor channel is represented as a query vector, augmented by a sensor-type embedding. Early fusion is conducted using cross-attention over the concatenated key/value pools, allowing missing modalities to be handled trivially by omission from the pool—no special masking heads are required (Zelic et al., 5 Apr 2026).
• Adapter Abstractions (Meta Engine): Each specialized backend (text, table, image) is wrapped in a uniform adapter interface, with a query router (statistical or learned) dispatching operators based on query type and historical performance (Li et al., 2 Feb 2026).
• Automatic Plan/Code Generalization (WorkflowLLM): Fine-tuning on a diverse, API-rich, hierarchical thought-annotated workflow dataset imparts strong generalization (e.g., to unseen APIs, out-of-distribution tasks), confirmed by gains in CodeBLEU and pass-rate metrics (Fan et al., 2024).

5. Performance Metrics, Comparative Evaluation, and Deployment

Unified models are evaluated by comparing end-to-end accuracy, efficiency, and robustness against traditional multi-component or specialized system baselines:

Selected empirical findings:

System	Domain	Unified vs. Baseline	Reference
DataFlow	LLM data prep	+3.7% Text2SQL exec acc., +7pts code	(Liang et al., 18 Dec 2025)
Meta Engine	Multimodal query	1.3–1.7× F1, up to 24× on outliers	(Li et al., 2 Feb 2026)
WorkflowLlama	Workflow orchestration	+9pt CodeBLEU, +9pt pass rate over GPT-4o (ICL)	(Fan et al., 2024)
PanLUNA	Biosignal FMs	81.21% bal-acc with 5.4M params (edges out 311M param baseline)	(Zelic et al., 5 Apr 2026)
UniRQR	Dialogue/Retrieval	SOTA F1, BLEU-2, and multi-task mutual enhancement	(Hu et al., 2024)

Practical implications:

• Model size reductions are typical, as a unified architecture can replace multiple specialized models.
• Latency and memory advantages are observed (e.g., PanLUNA’s INT8 deployment achieves 325.6 ms latency and 18.8 mJ per 10 s ECG inference on GAP9; unified LLM deployment obviates context switches between task-oriented sub-models).
• Robustness to partial inputs or dynamic context is enhanced, with missing modalities or incomplete specifications often handled without architectural modification (Zelic et al., 5 Apr 2026, Li et al., 2 Feb 2026).

6. Limitations, Trade-offs, and Future Directions

While unified query and workflow models offer significant abstraction, efficiency, and maintainability advances, they can entail complexity cost and specialization trade-offs:

• Complexity: The grammar/runtime (DashQL) or orchestration agent (DataFlow-Agent, WorkflowLlama) must bridge multiple domains, requiring substantial engineering for validation and debugging.
• Specialization: While unification yields cross-domain robustness and code/data reuse, certain optimizations (e.g., AM4 aggregation, modality-specific adapters) assume typical use-case patterns and may need extension for highly atypical modalities or tasks (Kohn et al., 2023, Li et al., 2 Feb 2026).
• Verification and Quality Control: Automated operator synthesis and plan generation (DataFlow-Agent, WorkflowLlama) raise the challenge of systematic verification and semantic correctness, although pipeline compilation, intent decomposition, and reference-based metrics (e.g., CodeBLEU, execution accuracy) address some of these concerns (Liang et al., 18 Dec 2025, Fan et al., 2024).
• Agentic Generalization: The emergence of agent-driven or prompt-driven pipeline assembly (WorkflowLLM, DataFlow-Agent) suggests future research in controllable, interpretable, and verifiable orchestration agents that operate against domain-general operator libraries.

A plausible implication is that, as abstraction barriers continue to dissolve, unified query and workflow models will become the foundation of agentic, data-centric, and multimodal automation frameworks spanning scientific, enterprise, and edge-device domains. This trend is evident in the systematic cross-domain improvements reported for unified architectures in biosignal intelligence (Zelic et al., 5 Apr 2026), process orchestration (Awad et al., 24 Jan 2025), LLM data preparation (Liang et al., 18 Dec 2025), and complex automation (Fan et al., 2024).