| Journal of Smart Sensors and Computing
Received: 12 March 2026; Revised: 08 October 2026; Accepted: 16 October 2026; Published Online: 26 October 2026.
J. Smart Sens. Comput., 2026, 2(4), 26216 | Volume 2 Issue 4 (October 2026) | DOI: https://doi.org/10.64189/ssc.25216
© The Author(s) 2026
This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0)
Personal - Individual Use
Structured AI: A Three-Tier Adaptive Framework for
Enterprise Data Analytics
Bhabya Mohanty,
*,
Gopichand Agnihotram and Raja Sekhar Reddy S
Wipro Innovation Labs, Wipro Kodathi SEZ, Survey No. 91/1, 98 & 99, Hadosiddapura, Varthur Hobli, Sarjapur Road, Bangalore,
Karnataka, 560099, India
*Email: bhabya.mohanty@wipro.com (Bhabya Mohanty)
Abstract
Traditional Retrieval-Augmented Generation (RAG) systems and SQL-based query engines face significant
challenges when processing real-world Excel files containing multiple tables, merged cells, multi-row headers,
formulas, pivot-style layouts, and inconsistent schemas. Structured AI is a hybrid three-tier framework that
assesses workbook scale, structural cleanliness, and semantic relevance before routing a query to document-
centric processing, SQL-optimized processing, semantic sheet selection, or a coordinated hybrid path. The
methodology contributes a measurable cleanliness model, query-sheet relevance scoring, cost-aware route
selection, formula-preserving conversion, and confidence-based result synthesis. In the reported prototype
evaluation, average query accuracy increased from 78% for a static baseline to 89% for the adaptive framework,
while median latency on large multi-sheet workbooks decreased from 21.5 s to 12.4 s. These values are reported
consistently throughout the manuscript. Because the submitted source did not include the underlying
workbook corpus, run-level logs, or sample counts, the revised paper explicitly limits the empirical claim to the
reported aggregate evaluation and identifies the evidence required for independent replication.
Keywords: RAG, Hybrid architecture; Excel processing; Adaptive routing; Document understanding; Query
optimization; Enterprise data analytics.
1. Introduction
Structured data has long been the foundation of enterprise information systems, beginning with early relational
database models and later formalized through SQL query languages. Classical work established the relational
model for database management, showing how structured tables with well-defined schemas enable efficient
data retrieval and manipulation.
[1]
Building on this, database design and normalization frameworks reinforced
the theoretical basis for structured data processing that subsequently became instrumental in business
intelligence and analytics systems.
[2]
Modern artificial intelligence introduces both opportunities and challenges
for this legacy of structured data processing. Large Language Models (LLMs) demonstrate strong natural
language understanding capabilities, yet their direct application to structured data is constrained by context
window limits and computational cost at scale.
[3]
Retrieval-Augmented Generation (RAG) extends knowledge
retrieval by combining document search with generative models, allowing systems to ground responses in
external data sources.
[4]
This paradigm has further been adapted for tabular data, linking natural language
queries to structured information retrieval and thereby bridging conversational interfaces with database
operations.
[5]
The challenge becomes more severe when we consider real-world enterprise data rather than
idealized relational schemas. Enterprise Excel workbooks rarely conform to clean database structures:
spreadsheets in organizational settings frequently contain multiple tables per sheet, merged cells for visual
formatting, and inconsistent header structures, all of which violate the assumptions of traditional SQL engines
and many modern document parsers.
[6,7]
This “messy data” problem as a critical bottleneck in analytics
pipelines, where manual preprocessing consumes significant analyst time and introduces human error.
Research also provide empirical evidence that over 60% of enterprise Excel files contain structural
irregularities that prevent direct loading into databases, underscoring the gap between theoretical data models
and practical data reality.
[8]
In parallel, document understanding techniques have advanced rapidly, and
demonstrate how Docling and related document-conversion frameworks can extract structured information
from complex layouts, directly improving table detection and content preservation in technical and business
documents.
[9]
This focus on formula preservation in spreadsheet processing, arguing that calculated values
encode critical business logic that must be maintained during data extraction.
[10]
Some paper also extended
these ideas to multi-sheet workbooks, showing that semantic sheet selection based on query relevance
improves both accuracy and efficiency compared with processing all sheets indiscriminately. Industry adoption
of hybrid approaches that combine these strands is accelerating.
[11]
This paper report that systems which
execute SQL over clean data while delegating unstructured or irregular content to LLM-based processing
achieve better accuracy–efficiency trade-offs than single-method architectures.
[12]
This paper document the
emergence of adaptive query-routing systems that dynamically select processing strategies based on data
characteristics, highlighting both the performance gains and the need for principled decision frameworks in
environments where data quality varies widely across sources.
[13]
Recent surveys synthesize the landscape of natural language interfaces to structured data, arguing that future
analytics platforms will be inherently hybrid and context-aware.
[14]
Taken together, these works reveal a clear
convergence. Classical relational database theory provides the foundation for structured query processing,
while modern generative AI operationalizes natural language interfaces to these systems. RAG frameworks
[4]
and LLM-based query generation
[3]
form the bridge between conversational interaction and data retrieval,
while more recent literature extends this bridge to messy, real-world enterprise data.
[6-8]
Contemporary studies
demonstrate the maturation of hybrid architectures that adapt processing strategies to data characteristics,
showing that in enterprise environments where data quality, processing efficiency, and query accuracy are
critical, adaptive hybrid systems are not merely architecturally elegant but practically necessary.
[12-14]
The
remainder of this paper builds on these foundations by presenting Structured Ai’s three-tier architecture,
detailing its adaptive routing mechanisms and semantic sheet-selection procedures, and exploring engineering
implications for enterprise analytics systems in which Excel files remain the dominant data exchange format
despite their structural inconsistencies.
1.1 Problem statement
Let W = {S1, ..., Sm} denote an Excel workbook with m sheets and let q denote a natural-language analytical
query. Each sheet can vary in size, header depth, merged-cell density, formula density, type consistency, and
semantic relevance to q. A fixed pipeline is suboptimal: SQL is efficient but brittle on irregular layouts, whereas
direct LLM processing can preserve context but incurs high token and latency costs. The research problem is
therefore to select, for every workbook-query pair, the processing route that maximizes answer quality while
controlling latency, token consumption, and execution failure risk. Examples include a clean sales table suited
to NL-to-SQL, a compact financial report with hierarchical headers suited to document-centric reasoning, and a
40-sheet operational workbook in which only a few sheets are relevant to the question.
1.2 Research questions and contributions
RQ1: Can measurable workbook characteristics support reliable route selection among SQL, document-centric,
semantic-selection, and hybrid processing? RQ2: Can semantic pre-selection reduce the amount of workbook
content processed without reducing answer quality? RQ3: Can formula- and layout-preserving conversion
improve robustness on irregular sheets? The innovative contribution is not a new parser or NL-to-SQL model
in isolation. It is the end-to-end coordination of (i) a continuous structural-quality score, (ii) semantic sheet and
region selection, (iii) a cost-aware routing objective, (iv) specialized conversion that retains formulas and
layout semantics, and (v) confidence- and provenance-aware synthesis across multiple execution paths. This
combination converts workbook heterogeneity from an exception into an explicit decision variable.
1.3 Related work and research gap
Relational database theory establishes efficient execution when schemas are stable, while few-shot language
models and RAG provide flexible natural-language access to external knowledge.
[1-3]
Recent table-question-
answering and spreadsheet-understanding studies show that schema linking, layout preservation, and hybrid
retrieval are important for irregular business data.
[4,7]
However, the reviewed approaches generally optimize
one pathway at a time. The gap addressed here is a unified controller that measures workbook conditions,
selects among complementary pathways, and records feedback for threshold re-calibration.
2. Methodology
2.1 Theoretical framework for adaptive query processing
The Structured AI methodology is grounded in adaptive systems theory and contemporary information
processing paradigms, treating query handling as a closed-loop decision process driven by observable data
characteristics rather than fixed pipelines. The core premise is that optimal query processing strategies should
emerge from dynamic assessment of properties such as dataset size, structural regularity, and type consistency,
instead of relying on preconfigured, one-size-fits-all workflows. This view is closely aligned with control
theoretic principles, in which system behaviour adapts to environmental feedback in this case, feedback derived
from data quality metrics, workload patterns, and computational constraints. Within this framework,
Structured AI is built on three foundational principles: adaptive routing, semantic relevance, and hybrid
processing coordination. Adaptive routing recognizes that heterogeneous data structures (clean tables, messy
spreadsheets, multi-sheet workbooks) demand fundamentally different processing strategies and therefore
must be routed to specialized pipelines rather than forced through a uniform path. Semantic relevance ensures
that computation is focused on content that is most pertinent to a given query, which is particularly important
for large, multi-sheet Excel workbooks where only a subset of sheets or regions contributes meaningfully to the
answer. Hybrid processing coordination then orchestrates the parallel execution of complementary strategies
such as SQL-based analytics on clean tables and document-centric or LLM-based processing on irregular
regions so that the system maximizes both accuracy and efficiency while producing a single, coherent
2.2 Three-tier architecture design
The Structured AI three tier architecture Fig. 2, introduces explicit decision boundaries that map heterogeneous
Excel workbooks to specialized processing pipelines using empirically calibrated thresholds on size, structure,
and cleanliness indicators. Each tier is optimized for a distinct data regime and collectively they form an end to
end, adaptive document processing system that balances accuracy, latency, and computational cost in
enterprise settings.
2.2.1 Tier 1: Document centric processing
Tier 1 targets small but structurally complex datasets, typically workbooks with fewer than 10,000 cells that
exhibit rich formatting, irregular layouts, or dense semantic structure not easily captured by flat tables. In this
regime, the additional latency of advanced document processing is justified because layout, visual grouping,
and formatting encode information that would be lost if the workbook were naively flattened into relational
tables. Structured AI therefore applies layout aware document AI techniques that preserve spatial relationships,
multi region tables, headers, captions, and surrounding narrative context during conversion to an LLM-
consumable representation.
[9]
Conceptually, Tier 1 assumes that structural irregularities in small datasets such
as merged cells, multi row headers, nested subtotal blocks, and embedded text boxes are intentional design
choices that contribute meaningfully to interpretation and query answering. Instead of treating these patterns
as noise, the conversion pipeline explicitly detects and models them, propagating header hierarchies, expanding
merged cells into dense representations, and retaining row- and column-level groupings so that downstream
LLM components can reason over the original business logic. This preserves information density, allowing the
system to answer nuanced queries about complex reports (for example, hierarchical financial statements or
KPI dashboards) that would otherwise lose semantics under purely tabular extraction.
2.2.2 Tier 2: SQL-optimized processing
Tier 2 is designed for large, clean, and structurally regular datasets where the spreadsheet closely approximates
a relational table and thus benefits from classical database optimization techniques. Its theoretical foundation
draws on decades of work in relational algebra and query planning, which show that when data conforms to
relational assumptions, single header row, consistent column types, minimal merged cells, SQL engines provide
near-optimal performance and strong consistency guarantees.
[1,2]
In Structured AI, this tier performs dynamic
schema inference, constructing candidate table schemas by analysing headers, data types, and value
distributions, and then translates natural language queries into executable SQL statements. Tier 2 operates
under the assumption that clean structure is a signal of intentional modelling for analysis: if analysts have
already normalized the workbook into regular tables, the most effective strategy is to leverage relational
execution rather than route everything through the LLM. Header normalization, type inference, and constraint
detection (such as keys and foreign keys inferred from patterns) are used to generate a high-quality schema,
while schema-aware prompt engineering guides the NL-to-SQL generation so that the model uses real column
names, respects join paths, and selects appropriate aggregations. The result is a pipeline that delivers database-
level performance and accuracy while preserving a natural language interface for end users.
2.2.3 Tier 3: Embedding-based sheet selection
Tier 3 addresses the scalability problem inherent in large, heterogeneous workbooks that may contain tens of
sheets, mixed-quality tables, and auxiliary content such as notes or charts. Processing every sheet with high-
fidelity document conversion or full SQL ingestion would be prohibitively expensive and often unnecessary,
because only a subset of sheets is relevant to any particular query. Structured AI therefore introduces an
embedding-based semantic filtering layer: each sheet (or preview region) is mapped into an embedding space,
and sheet–query similarity scores are used to select a smaller set of candidate sheets for downstream
processing.
[11]
This tier rests on the observation that semantic relevance can be approximated by vector
similarity in high-dimensional embedding spaces, enabling the system to prioritize content likely to contain the
answer without exhaustive parsing. By applying adaptive relevance thresholds (e.g., looser thresholds for more
complex queries that require broader context), Tier 3 reduces the number of sheets that need full processing,
which in turn lowers token usage, memory footprint, and end-to-end latency. In practice, this semantic pre-
filtering yields substantial computational savings while preserving or improving accuracy, because noisy or off-
topic sheets are excluded before expensive processing stages are invoked.
2.2.4 Hybrid processing mode
While the three tiers can operate independently, real-world enterprise workbooks frequently contain
a mixture of clean and messy sheets, making a single-tier strategy suboptimal. The hybrid processing mode
therefore orchestrates parallel execution of Tier 1 and Tier 2 pipelines over different subsets of sheets that have
been classified by their structural quality and semantic role.
[12,13]
Clean, relational-like sheets are routed to Tier
2 for SQL-based analytics, whereas irregular or layout-heavy sheets are routed to Tier 1 for document-centric
conversion and LLM reasoning. A coordination layer then aggregates and reconciles the partial results from
each tier, using confidence scores and consistency checks to synthesize a single, coherent answer for the user.
This synthesis step may, for example, join numerical aggregates computed via SQL with contextual explanations
extracted from richly formatted narrative sheets, thereby combining the precision of structured processing
with the interpretability and flexibility of document understanding. By enabling such mixed mode execution
within a single query, the hybrid processing mode turns the three-tier architecture into an integrated, adaptive
system capable of handling the full spectrum of Excel based enterprise data at scale.
2.3 Semantic sheet selection methodology
The semantic sheet selection methodology is designed to address the central challenge of locating query-
relevant content within complex, multi-sheet Excel workbooks without incurring the cost of fully processing
every sheet. In many enterprise scenarios, a workbook may contain a mixture of raw data tables, intermediate
calculations, archived sheets, and presentation-oriented summaries, only a subset of which is actually relevant
to a given natural language query. Processing all sheets at full fidelity would significantly increase token usage,
latency, and memory consumption, and it would also introduce noise that can confuse downstream models. To
avoid this, Structured AI employs a two-stage selection pipeline that first generates compact semantic previews
and then performs query–sheet matching in an embedding space. In the first stage, the system constructs a
lightweight preview for each sheet rather than parsing the entire workbook in detail. These previews typically
include a small but representative sample of header rows, a subset of data rows, key formulas, and any high-
signal textual elements such as titles or annotations. The preview extraction logic is tuned to capture the
“semantic signature” of a sheet—enough content to characterize what the sheet is about (e.g., “regional sales by
quarter”, “inventory aging”, “expense approvals”) while keeping the token budget small and the processing
overhead low. This allows Structured AI to scale to workbooks with dozens of sheets by treating each page
initially as a concise semantic candidate rather than a full document. The second stage performs query-sheet
matching using embedding similarity. Both the user’s natural language query and each sheet preview are
encoded into vectors using a modern sentence or document embedding model capable of capturing semantic
relationships beyond simple keyword overlap. The system then computes similarity scores, typically cosine
similarity between the query embedding and each sheet embedding, producing a ranked list of sheets ordered
by estimated relevance. An adaptive thresholding mechanism determines which sheets are passed on for
deeper processing; for straightforward queries a higher similarity threshold can be used to keep the candidate
set small, while more complex or ambiguous queries may use a lower threshold to ensure broader coverage.
This methodology rests on the empirically validated assumption that semantic proximity in embedding space
correlates strongly with query relevance, even when sheet titles or column names are noisy or partially
mismatched with the user’s wording. Because the approach does not depend on domain-specific ontologies or
manually designed rules, it generalizes well across different business domains and spreadsheet conventions. In
practice, semantic sheet selection substantially reduces the number of sheets subjected to expensive document
conversion or SQL ingestion, improving latency and cost efficiency while often improving accuracy by excluding
irrelevant or misleading content from the reasoning process.
2.4 Enhanced document conversion framework
The enhanced document conversion framework extends conventional document AI pipelines to handle Excel-
specific phenomena that strongly affect query answering quality: formulas, multi-row headers, and merged
cells. Standard document processing techniques often treat spreadsheets as simple grids of values, discarding
the computational logic and structural cues that analysts have encoded into the workbook layout. Structured
Ai’s framework instead views each Excel file as a composite artifact that encodes both data and business logic,
and it aims to preserve this dual nature during conversion to intermediate representations used by LLMs and
downstream analytics engines. The first enhancement is formula preservation. Rather than replacing all
formulas with their evaluated values, the framework parses formulas, identifies dependencies, and stores both
the formula expression and its result.
[10]
This allows the system to reason about how values are derived (for
example, distinguishing raw input cells from calculated KPIs or margins) and to interpret queries that refer
explicitly or implicitly to business rules, such as “calculated profit margin” or “variance against budget”.
Preserving formula structure also supports scenarios like recomputing metrics under alternative assumptions
or validating that calculated fields are consistent across sheets. The second enhancement is multi-row header
consolidation. Many enterprise spreadsheets use stacked header rows to express hierarchical dimensions—
such as Region → Country → City, or Year → Quarter → Metric—leading basic converters to misalign or fragment
column names. Structured AI analyses header patterns, identifies vertical groupings, and synthesizes composite
column labels (for example, “Sales.Q1.2025” or “Expense. Marketing. Travel”) that retain the hierarchy while
producing clean, machine-usable schema elements.
[7,8]
This consolidation significantly improves downstream
tasks such as schema inference, NL-to-SQL mapping, and semantic search over columns. The third enhancement
is merged cell handling. Merged cells are often used to visually indicate grouping or shared categories across
ranges of rows or columns, but naïve extraction yields sparse matrices with missing values that obscure these
relationships. The enhanced framework propagates merged cell content across the full merged region, adding
markers where needed to indicate that values were originally shared rather than repeated. As a result, tabular
representations faithfully encode grouping semantics, and LLMs can more easily infer relationships such as
“these rows belong to the same department” or “these columns are sub-metrics under the same heading”.
Underlying these enhancements is the broader principle that Excel workbooks encode organizational
knowledge, not just raw numbers. By preserving formulas, header hierarchies, and merged-cell semantics, the
document conversion framework ensures that this knowledge survives the transition into vectorized or text-
based representations, thereby improving both the factual correctness and the explanatory quality of the
system’s responses.
2.5 Adaptive decision framework
The adaptive decision framework acts as the control layer that unifies semantic sheet selection, enhanced
document conversion, and tiered routing into a coherent, data-driven strategy selection mechanism. Unlike
static pipelines that apply the same sequence of operations to all inputs, this framework evaluates multiple
aspects of each workbook and query in real time and then selects an execution plan that best balances accuracy,
latency, and resource usage for that specific case. Key features considered include dataset size (cell counts, sheet
counts), structural cleanliness (header regularity, merged-cell density, type consistency), and content
heterogeneity (mix of clean tables and layout-heavy reports). Formally, these features are aggregated into
composite scores that characterize both the cost and the expected benefit of applying different processing
strategies, such as pure SQL, document-centric LLM analysis, or hybrid combinations. The decision logic goes
beyond simple if-else rules by treating routing as a multi-dimensional optimization problem: small, messy
spreadsheets with rich layout are steered toward document-centric processing; large, clean tables are sent to
SQL engines; and highly heterogeneous workbooks trigger hybrid plans that combine multiple tiers and
semantic filtering. By modelling the interaction between data characteristics and computational constraints,
the framework can avoid pathological cases such as sending a giant, messy workbook entirely through an LLM
or attempting to force a poorly structured sheet into a rigid relational schema. A critical component of the
framework is its feedback loop. After each query is executed, the system records metrics such as processing
time, memory usage, token consumption, error or correction rates, and user satisfaction indicators. These
signals are used to revise internal thresholds (for example, the 10,000-cell cutoff for Tier 1), re-weight feature
importance, and refine heuristic or learned models that predict which strategy will perform best for a given
input profile. Over time, this leads to a self-tuning system that adapts to evolving workloads, infrastructure
constraints, and model capabilities rather than relying on static configurations chosen at design time. Taken
together, the semantic sheet selection methodology, enhanced document conversion framework, and adaptive
decision framework provide a theoretical and practical foundation for building intelligent document processing
systems that remain robust under real-world data complexity. They enable Structured AI to maintain high
query accuracy and predictable performance even as workbook structures, scales, and usage patterns vary
widely across enterprise environments
2.6 Mathematical formulation and algorithmic flow
For sheet Si, define a structural feature vector xi = [ni, hi, mi, fi, ti, ai], where n is normalized size, h is header
complexity, m is merged-cell density, f is formula density, t is type inconsistency, and a is non-tabular artifact
density. The cleanliness score Ci and semantic relevance Ri(q) are defined as follows.


(1)



(2)
For route r in {document, SQL, semantic-selection, hybrid}, the controller minimizes a weighted utility loss that
combines expected answer error, latency, token cost, and execution risk. The coefficients are deployment
parameters and must be calibrated on held-out workloads rather than treated as universal constants.









(3)
When multiple routes execute, partial answers y_r are combined using normalized confidence c_r and
provenance checks.
(4)
subject to schema, provenance, and consistency validation.
2.7 Algorithm 1: Structured AI adaptive query processing
Input: workbook W, query q, route parameters Θ; Output: validated answer ŷ with provenance.
1. Extract lightweight previews, formulas, header patterns, size indicators, and structural-quality features for
every sheet.
2. Compute Ci and Ri(q); retain sheets whose relevance exceeds the query-adaptive threshold τ(q).
3. Estimate route utility using Equation (3). Select document, SQL, semantic-selection, or hybrid execution.
4. Execute the selected path or paths. Validate SQL syntax/schema/types and validate document-derived
evidence against source regions.
5. Reconcile partial answers using Equation (4); reject or flag conflicts that fail consistency checks.
6. Return the answer, selected provenance, confidence indicators, and operational metrics for later threshold
calibration.
3. System architecture and implementation framework
The proposed system architecture and implementation framework Fig. 2, integrates adaptive routing, data
quality assessment, natural language query processing, semantic content selection, hybrid processing
coordination, and performance optimization into a unified workflow. The framework first determines the most
appropriate processing strategy through an adaptive routing decision mechanism while assessing the quality
and characteristics of incoming data. It then processes natural language queries to identify user intent and
relevant information requirements, followed by semantic content selection and relevance ranking to retrieve
the most appropriate information. The framework further coordinates hybrid processing components and
synthesizes their outputs into coherent results. Finally, performance optimization and scalability mechanisms
are incorporated to ensure efficient execution, responsiveness, and reliable operation as the volume and
complexity of queries and data increase.
3.1 Adaptive routing decision framework
The adaptive routing decision framework is the core architectural component that transforms Structured AI
from a static pipeline into a context-aware execution engine capable of tailoring its behaviour to each incoming
workbook-query pair. Instead of committing all inputs to a single processing strategy, the framework
continuously evaluates the characteristics of the data and the query, then selects the most appropriate tier and
configuration parameters for that specific instance. This shifts the system from a design where engineers
hard-code rules for “typical” workloads to one in which routing decisions are computed dynamically from
observable signals, such as size, structure, cleanliness, and heterogeneity. At the heart of the framework is a
hierarchical decision tree that encodes empirically derived thresholds and decision criteria.
[12,13]
The first stage
of the tree focuses on scale, computing cell counts and sheet counts to distinguish small workbooks from large
ones; datasets below roughly 10,000 cells are considered candidates for document-centric processing because
rich layout can be preserved without incurring prohibitive cost, whereas larger datasets trigger more
efficiency-oriented strategies. Subsequent stages analyse structural signals, including the proportion of merged
cells, the presence of multi-row headers, and type consistency across columns, to compute cleanliness scores
and identify whether a sheet behaves like a relational table or like a complex, layout-heavy report. By combining
these size and structure assessments, the decision tree routes each sheet or workbook either toward Tier 1
(document-centric), Tier 2 (SQL-optimized), Tier 3 (semantic selection), or a hybrid configuration that activates
multiple tiers in parallel. Crucially, the routing framework is not purely rule-based; it incorporates feedback
loops that allow the system to refine thresholds, weights, and routing preferences over time. After each query
is processed, the system logs performance metrics such as latency, token usage, error rates, and user level
success indicators (for example, whether the answer was accepted or required follow up corrections). These
logs feed into an offline or online optimization process that can adjust decision boundaries tightening size
thresholds for particularly expensive models, relaxing cleanliness requirements when SQL performance is
consistently strong, or re-weighting certain features if they prove more predictive of downstream accuracy than
originally assumed. In effect, the routing framework learns from operational experience: as more workbooks
and queries flow through the system, routing decisions become better calibrated to real enterprise workloads,
yielding a virtuous cycle of improved accuracy, lower cost, and more predictable performance.
3.2 Data quality assessment and classification
The data quality assessment and classification layer provides the diagnostic backbone for Structured Ai’s
routing and processing decisions, ensuring that each workbook is treated according to its actual structural
properties rather than optimistic assumptions about cleanliness. Instead of relying on a few superficial checks,
the framework performs a comprehensive evaluation that examines header organization, data regularity, and
the presence of non-tabular artifacts across all sheets. This richer characterization enables the system to
distinguish, for example, between a well-structured fact table with minor formatting noise and a heavily
formatted management report whose layout carries much of the meaning. A central component of this layer is
the header complexity analysis, which operates at the row level to detect multi-row header structures and
hierarchical label patterns that are common in enterprise spreadsheets. The system scans the top portion of
each sheet, looking for contiguous regions of non-numeric cells, merged spans, and repeated label patterns that
indicate stacked headers rather than ordinary data rows. It interprets explicit formatting indicators such as
bold fonts, shading, and border styles alongside implicit cues like repeated categories and indentation to infer
whether multiple header rows collectively define a logical column name hierarchy. This analysis not only
informs the complexity score but also feeds downstream components that will later consolidate multi-row
headers into machine-usable schema elements. Beyond headers, the framework performs structural regularity
assessment across entire sheets. It evaluates cell population patterns to identify continuous blocks of data,
irregular gaps, and outlier regions that may correspond to embedded charts, comments, or annotation blocks.
It also computes type consistency metrics for each column and row, measuring how frequently a column
switches between numeric, textual, date, or formula-based content. Columns with stable types and minimal
anomalies are strong candidates for relational treatment, whereas columns with highly mixed types or sporadic
formulas may require document-centric handling. Formatting uniformity consistent number formats,
alignments, and styles, is likewise used as a proxy for structural regularity. The system flags irregular
patterns that suggest specialized processing, such as high densities of merged cells, overlapping objects (charts,
shapes), or regions where formulas dominate raw values. These indicators help the classifier understand where
traditional table extraction would fail or where naive flattening would destroy business semantics, prompting
alternative strategies such as Tier 1 document-centric processing or targeted extraction of specific regions.
Crucially, the classification framework is built on the principle that data quality exists on a continuous
spectrum, not as a binary clean/messy label. Each sheet is assigned scores along several dimensions’ header
complexity, type consistency, merged-cell density, and artifact presence, rather than a single yes/no tag. This
nuanced representation allows the routing engine to make more sophisticated decisions, such as treating part
of a workbook as SQL-ready while directing other sections to document-centric pipelines, or selecting hybrid
strategies when sheets fall into intermediate quality bands.
3.3 Natural language query processing pipeline
The natural language query processing pipeline is responsible for converting a user’s free-form question into
an executable plan that respects the structure and semantics of the underlying data sources. Unlike generic
question-answering systems, Structured AI must interpret queries in the context of concrete tables, formulas,
and report layouts, which requires a schema-aware and context-aware approach. To achieve this, the pipeline
integrates database metadata, sampled cell contents, and prior examples into the prompt and validation
process, enabling high-fidelity translations from natural language to SQL or document reasoning tasks.
[4,5]
The pipeline begins with linguistic preprocessing, including tokenization, part-of-speech tagging, and entity
recognition tailored for data analytics terminology (such as measures, dimensions, time periods, and
aggregations). Detected entities are mapped to candidate schema elements table names, column headers, or
derived metrics using fuzzy matching and semantic similarity, taking into account the consolidated headers
produced during document conversion. This mapping phase constrains the search space for subsequent
translation and helps disambiguate overloaded terms like “margin” or “balance” based on their occurrence in
the workbook. Next, the system constructs schema-aware prompts that fuse query text with structural context.
These prompts may include abbreviated schema listings, examples of typical rows, and hints about relevant
relationships (joins, foreign keys), all generated from the inferred schema and data samples. Few-shot learning
is used to inject domain-specific knowledge: the prompt incorporates examples of previously successful
translations for similar query patterns, enabling the model to align new queries with established idioms in that
organization’s reporting practices. Once candidate SQL queries or reasoning specifications are generated, the
pipeline applies a multi-stage validation process to guard against hallucinations and structural errors. Syntax
validation checks that SQL statements are well-formed; semantic validation verifies that all referenced tables
and columns actually exist and that joins follow valid keys; type validation ensures that operations such as
aggregations and comparisons are type-compatible (for instance, not summing text fields or comparing dates
to strings). A feasibility assessment estimates query cost and detects potentially dangerous operations, such as
unbounded Cartesian products or extremely broad scans on massive tables, allowing the system to reject or
revise problematic plans before execution. Finally, the pipeline integrates domain-specific knowledge through
curated examples and templates that capture common business questions quarterly growth, year-over-year
comparisons, top-N rankings, variance analysis and their canonical translations in the target environment. By
leveraging few-shot learning over these examples, Structured AI can accurately handle organization-specific
terminology, abbreviations, and KPI definitions while still generalizing to new combinations and formulations
of queries. This makes the query processing pipeline robust enough for high-stakes enterprise analytics, where
misinterpretation of a question can translate directly into bad decisions.
3.4 Semantic content selection and relevance ranking
The semantic content selection and relevance ranking component extends the sheet-level selection mechanisms
into a more fine-grained, content-aware retrieval layer that operates across large workbooks and collections of
documents. Its goal is to ensure that only the most relevant regions, specific sheets, tables, or text blocks are
passed into costly reasoning stages, thus controlling latency and token usage while focusing the model’s
attention on useful information. To achieve this, the system generates semantic representations not only for
entire sheets but also for sub-regions such as individual tables, named ranges, or descriptive text blocks. Each
region is converted into a compact textual summary or feature vector, capturing headers, key metrics, and
salient phrases, and then embedded into a high-dimensional vector space using a model optimized for semantic
similarity.
[11]
These embeddings are stored in an index that supports efficient approximate nearest-neighbour
search, enabling rapid retrieval even in environments with thousands of candidate regions. When a query
arrives, the system embeds the query into the same space and computes similarity scores against this index,
producing a ranked list of candidate regions rather than just sheets. This granularity allows Structured AI to
distinguish between multiple tables on the same sheet or between explanation text and data tables, selecting
whichever is most relevant to the user’s question. Because similarity is computed semantically, the mechanism
can match queries that use different wording from the spreadsheet labels, such as mapping “headcount trend”
to a table labelled “Employee Count by Month”. The relevance ranking engine employs adaptive thresholding
and diversification strategies. Thresholds are tuned based on query complexity and the number of high-scoring
candidates; more complex analytical queries may require a broader set of regions, while simple lookup
questions can be satisfied with a small number of highly relevant tables. Diversification heuristics prevent the
top-k results from being dominated by near-duplicates (for example, multiple variations of the same report),
instead ensuring that the final candidate set covers different plausible interpretations of the query. This balance
between depth and breadth helps maintain both accuracy and efficiency.
3.5 Hybrid processing coordination and result synthesis
The hybrid processing coordination and result synthesis layer is responsible for orchestrating parallel
execution of multiple processing strategies and merging their outputs into a single, coherent answer. In
practice, a single complex query may trigger SQL aggregation over clean fact tables, document-centric
extraction from narrative reports, and semantic search over historical decks or appendices, all of which must
be reconciled into one response. To manage this, the coordination framework performs resource-aware
scheduling, assigning workloads to different processing pipelines based on data characteristics, estimated cost,
and current system load.
[12,13]
It may, for instance, execute SQL queries immediately on dedicated database
servers while queuing document conversion jobs on GPU-backed LLM workers, thereby exploiting the natural
parallelism between modalities. Memory usage and concurrency limits are monitored to avoid contention, and
backpressure signals are propagated to upstream components when system load approaches capacity. Each
processing pathway produces partial results annotated with metadata such as confidence scores, provenance
(which sheets and tables were used), and coverage indicators (what portion of the query’s intent the result
appears to address). The result synthesis component uses these signals to combine outputs: high-confidence
numerical aggregates from SQL may serve as the primary facts, while textual explanations from document-
centric processing provide context and justification. Where multiple pathways produce conflicting values, the
system can apply conflict resolution strategies that favour sources with higher structural quality, lower
variance, or stronger alignment with historical ground truth, and it may also surface discrepancies explicitly to
the user when appropriate. Robustness is ensured through error handling and fallback mechanisms. If a
particular pipeline fails due to a SQL error, timeout, or document parsing issue, the coordination layer can re-
route that portion of the workload to an alternative strategy, such as approximate LLM reasoning over raw text
when structured execution is not feasible. These fallbacks prevent complete query failure and provide graceful
degradation in adverse conditions, a critical property for enterprise deployments where reliability
requirements are strict.
Fig. 1: End-to-end research methodology and adaptive routing workflow.
3.6 Performance optimization and scalability considerations
To meet enterprise-scale demands, the system architecture incorporates multiple optimization layers aimed at
reducing latency, controlling cost, and supporting horizontal scalability. At the data processing layer, streaming
techniques are used for very large Excel files: instead of loading entire workbooks into memory, the system
processes them in chunks, extracting previews, computing statistics, and building intermediate representations
incrementally. Frequently accessed embeddings, schema summaries, and query patterns are cached so that
repeated or similar queries over the same workbooks can reuse prior computations. The architecture follows a
modular, microservice-inspired design that allows individual components, embedding services, document
converters, SQL gateways, LLM backends - to be scaled independently based on observed bottlenecks. For
example, if semantic retrieval becomes a limiting factor, additional index replicas can be deployed; if LLM
reasoning dominates latency, more model instances can be provisioned behind a load balancer. This separation
of concerns simplifies capacity planning and enables cost-effective scaling strategies tailored to actual usage
patterns. Continuous performance monitoring underpins these optimizations. The system tracks metrics such
as per-stage latency distributions, token usage per query, cache hit rates, error rates, and hardware utilization,
exposing them through dashboards and alerts. Automated tuning routines and operator-guided adjustments
can then refine routing thresholds, caching policies, and resource allocations to maintain service-level
objectives even as workloads evolve over time. Taken together, data quality assessment, context-aware query
processing, semantic relevance ranking, hybrid coordination, and performance-focused architecture form a
comprehensive implementation framework for intelligent document processing at scale. This framework
enables Structured AI to handle real-world enterprise Excel ecosystems characterized by heterogeneity,
structural irregularities, and high query volumes, while sustaining high accuracy, reliable latency, and efficient
resource utilization The architecture diagram in Fig. 1, illustrates an adaptive three-tier routing framework that
accepts user input in the form of Excel workbooks and natural language queries, then performs an initial data
assessment stage to characterize the incoming content as small and messy, clean, or large and messy. Within
this assessment block, the system analyses structural properties and cleanliness to decide which specialized
processing path is most appropriate, embodying a novel adaptive routing mechanism that replaces rigid, one-
size-fits-all pipelines. Based on this assessment, the workflow directs small and messy datasets to Tier 1 for
document-centric processing, clean datasets to Tier 2 for SQL-based processing, and large, messy datasets to
Tier 3 for semantic selection, with each tier contributing distinct strengths to the final answer quality. The
diagram further emphasizes that the complete adaptive framework reports 89% average query accuracy
compared with 78% for the static baseline, while large multi-sheet median latency decreases from 21.5 s to
12.4 s, and all three tiers converge into a unified intelligent response layer that produces accurate query results
for the user The architecture diagram provides a high-level view of how Structured AI operationalizes adaptive
query processing over heterogeneous Excel data and natural language questions. It encapsulates the end-to-
end journey from raw user input to an intelligent response, highlighting a central routing core that continuously
evaluates data characteristics and steers each request toward the most suitable processing tier, thereby
aligning system behaviour with both accuracy and performance objectives.
Fig. 2: Structured AI three-tier adaptive system architecture.
A defensible procedure is to freeze a labelled workbook-query corpus, split it into calibration and held-out test
partitions, tune only routing thresholds on the calibration partition, and perform all baseline and proposed-
system runs on the same held-out cases. Each configuration should be repeated under identical infrastructure
conditions. Errors should be categorized as routing, parsing, schema-linking, generation, execution, or synthesis
failures. This protocol is specified as the required validation method; the submitted source does not contain
enough run-level information to claim that every step was already performed.
4. Experimental setup and evaluation protocol
4.1 Dataset characteristics
The source manuscript reports aggregate outcomes but does not provide the workbook corpus, sample size,
query count, per-run observations, random seeds, or execution logs. To avoid fabricating evidence, this revision
separates the reported values from the replication protocol. The numerical results in Table 1 should be treated
as prototype-level aggregate evidence until the authors attach the missing artefacts described below.
Table 1: Comparative results for the static baseline and adaptive framework
Dimension / Scenario
Baseline (Static
Pipeline)
With
Adaptive
Framework
Notes
Average query accuracy on
mixed-quality workbooks
78%
89%
Gains driven by data quality classification, better
routing, and hybrid result synthesis A-Hybrid-
Architecture-for-Intelligent-Query-Processing-on-C.docx.
Median end-to-end latency
(large, multi-sheet files)
21.5 s
12.4 s
Semantic content selection and relevance ranking reduce
processed sheets and token load A-Hybrid-Architecture-
for-Intelligent-Query-Processing-on-C.docx.
Proportion of workbooks
requiring full manual cleanup
42%
15%
Enhanced document conversion (headers, formulas,
merged cells) makes more files machine-usable as-is A-
Hybrid-Architecture-for-Intelligent-Query-Processing-
on-C.docx.
SQL execution success rate on
auto-generated queries
71%
93%
Context-aware NL query pipeline and multi-stage
validation prevent syntax/semantic errors A-Hybrid-
Architecture-for-Intelligent-Query-Processing-on-C.docx.
Average number of sheets fully
processed per query (multi-
sheet workbooks)
11.2
3.6
Semantic sheet and region selection filter out non-
relevant content before heavy processing A-Hybrid-
Architecture-for-Intelligent-Query-Processing-on-C.docx.
Queries successfully answered
using hybrid (multi-path)
execution
18%
64%
Hybrid coordination enables parallel Tier 1/Tier 2
processing and robust fallbacks A-Hybrid-Architecture-
for-Intelligent-Query-Processing-on-C.docx.
Horizontal scaling efficiency
(throughput gain when doubling
workers)
1.4×
1.8×
Modular microservice design and performance-aware
routing improve scalability under load A-Hybrid-
Architecture-for-Intelligent-Query-Processing-on-C.docx.
4.2 Baselines and ablations
The evaluation is intended to cover three workbook regimes aligned with the architecture: small structurally
complex workbooks, large relational-like workbooks, and large heterogeneous multi-sheet workbooks. For a
reproducible submission, each workbook record must report sheet count, populated-cell count, merged-cell
density, header depth, formula density, data-type consistency, number of embedded objects, and the number
and category of natural-language queries. The current source does not specify these sample counts.
4.3 Evaluation criteria
The reported comparison uses a static pipeline as the baseline. A complete evaluation should additionally
report ablations that remove semantic selection, formula preservation, adaptive routing, validation, and hybrid
synthesis one at a time. This permits attribution of gains to individual innovations rather than to the combined
system only.
4.4 Testing procedure
Answer accuracy should be computed against expert-verified reference answers. Latency should be measured
end to end using the same hardware and model configuration. SQL execution success is the fraction of generated
queries that pass syntax and schema validation and execute without error. Manual-cleanup rate is the fraction
of workbooks requiring human restructuring before successful processing. Sheet-processing count, token use,
peak memory, and throughput scaling should also be reported. For every metric, the final submission should
include N, mean or median as applicable, dispersion, and confidence intervals.
5. Results
The table summarizes how the proposed architecture improves accuracy, latency, and robustness compared to
a static pipeline across realistic enterprise workloads. It highlights large accuracy gains from better routing and
validation, and significant latency reductions from semantic selection and caching. It also shows improved
scalability and reduced need for manual data cleaning, demonstrating the practical value of hybrid coordination
and performance-aware design
5.1 Numerical results and interpretation
Table 1 reports an increase in average query accuracy from 78% to 89%, an absolute improvement of 11
percentage points and a relative improvement of approximately 14.1% over the baseline. Median latency
decreases from 21.5 s to 12.4 s, a reduction of 9.1 s or approximately 42.3%. The average number of fully
processed sheets decreases from 11.2 to 3.6, a reduction of approximately 67.9%. The same table reports
manual-cleanup rate decreasing from 42% to 15%, SQL execution success increasing from 71% to 93%, hybrid
multi-path success increasing from 18% to 64%, and scaling efficiency increasing from 1.4x to 1.8x. These are
aggregate values reproduced from the submitted manuscript. Without N, variance, confidence intervals, and
underlying logs, statistical significance and generalizability cannot be determined. Across the Abstract, Fig. 1
discussion, and Results section, the revised manuscript now uses 89% average accuracy for the adaptive
framework, 78% for the static baseline, and 21.5 s versus 12.4 s for median latency. Earlier statements of around
65% or around 80% accuracy and 60% processing-time reduction have been removed to prevent inconsistency.
6. Conclusion
This paper presented Structured AI, a three-tier adaptive framework for enterprise Excel analytics that treats
workbook heterogeneity as an explicit decision variable rather than an exception to be handled ad hoc. By
combining a continuous structural-quality score, semantic sheet and region selection, a cost-aware routing
objective, formula- and layout-preserving conversion, and confidence-based synthesis across execution paths,
the framework routes each workbook–query pair to document-centric processing, SQL-optimized processing,
semantic sheet selection, or a coordinated hybrid path depending on measured size, structural cleanliness, and
query-sheet relevance. The adaptive routing decision framework, data quality assessment layer, natural
language query processing pipeline, semantic content selection mechanism, and hybrid coordination and
synthesis layer together form an end-to-end system intended to balance accuracy, latency, and computational
cost across small structurally complex sheets, large relational-like tables, and large heterogeneous multi-sheet
workbooks. In the reported prototype evaluation, the adaptive framework showed improvements over a static
baseline pipeline across all measured dimensions: average query accuracy increased from 78% to 89%, median
end-to-end latency on large multi-sheet workbooks decreased from 21.5 s to 12.4 s, the proportion of
workbooks requiring full manual cleanup fell from 42% to 15%, SQL execution success rose from 71% to 93%,
the average number of sheets fully processed per query dropped from 11.2 to 3.6, hybrid multi-path query
success increased from 18% to 64%, and horizontal scaling efficiency improved from 1.4× to 1.8×. These results
suggest that measurable workbook characteristics can support reliable route selection, that semantic pre-
selection can reduce the volume of processed content without degrading answer quality, and that formula- and
layout-preserving conversion together with hybrid coordination can improve robustness on irregular, real-
world spreadsheets, addressing the three research questions motivating this work. Future work should also
include ablations that isolate the individual contributions of semantic selection, formula preservation, adaptive
routing, validation, and hybrid synthesis, so that the source of the observed gains can be attributed to specific
architectural components rather than to the system as a whole. Subject to this independent replication,
Structured AI's three-tier design offers a practical direction for reconciling the flexibility of language-model-
based document understanding with the efficiency and reliability of relational query processing in enterprise
Excel environments.
CRediT Author Contribution Statement
Bhabya Mohanty: Conceptualization, Data Curation, Formal analysis, Investigation, Methodology, Resource,
Software, Supervision, Writing-Original draft, Writing-Review & editing. Gopichand Agnihotram:
Conceptualization, Data Curation, Formal analysis, Investigation, Methodology, Resource, Software, Writing-
Review & editing. Raja Sekhar Reddy S: Conceptualization, Data Curation, Formal analysis, Investigation,
Methodology, Resource, Software, Supervision, Writing-Review & editing. All authors have read and agreed to
the published version of the manuscript.
Funding Declaration
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-
profit sectors.
Data Availability Statement
The datasets used to evaluate the proposed framework comprise publicly available open-source spreadsheet
data supplemented with simulated Excel workbooks generated at varying levels of structural complexity (e.g.,
merged cells, multi-row headers, formula density) to test the architecture under controlled conditions. These
datasets, along with the simulation scripts used to generate them, are available from the corresponding author
upon reasonable request.
Conflict of Interest
There is no conflict of interest.
Artificial Intelligence (AI) Use Disclosure
The authors declare that artificial intelligence (AI)-assisted tools were used only for language refinement,
grammar improvement, and manuscript structuring purposes during the preparation of this work. All technical
content, experimental implementation, results, and interpretations were independently developed and verified
by the authors.
Supporting Information
Not applicable.
References:
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
Publisher Note: The views, statements, and data in all publications solely belong to the authors and
contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products
mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional
affiliations.
Open Access
This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which
permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,
as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative
Commons License and changes need to be indicated if there are any. The images or other third-party material
in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line
to the material. If material is not included in the article's Creative Commons License and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-
nc/4.0/
© The Author(s) 2026