From SQL Generation to Tool Selection: A Domain-Oriented Pattern for MCP Servers

Bartolomeo Bogliolo

cs.AIAug 22, 2026

From SQL Generation to Tool Selection: A Domain-Oriented Pattern for MCP Servers

Authors: Bartolomeo Bogliolo

Abstract

Agents built on Large Language Models (LLMs) increasingly reach enterprise data through the Model Context Protocol (MCP), and many MCP database servers maximize flexibility by exposing a single generic SQL execution tool. This paper proposes the Domain-Oriented Tooling Pattern: instead of generating SQL at query time, the model selects from a small set of domain-aligned tools whose parameterized queries encapsulate schema navigation, joins and business rules on the server side. We formalize the pattern around three architectural invariants and introduce Model Demotion, the observation that replacing SQL synthesis with intent classification lowers the model tier required to serve routine requests. As a reference implementation we present MCP Blueprint, an open-source framework in which domain tools are defined declaratively as YAML metadata plus external parameterized SQL files. We evaluate the pattern with a public reproducibility benchmark comparing three MCP server designs - raw SQL execution, a thin generic tool pack, and a verticalized domain pack - on four local models (3B-8B) across seventeen customer-facing tasks over the Sakila database (609 completed cells; temperature 0; three repetitions per cell). The verticalized pack reaches a pooled mean score of 0.939 versus 0.666 for raw SQL and 0.605 for the generic pack; the smallest model improves from 0.583 to 0.929, matching or exceeding every larger configuration while cutting cost per correct answer by an order of magnitude. All harness code, prompts, gold answers, frozen packs and per-cell results are publicly available.

Figures & tables

Dimension	Generic SQL interface	Domain-oriented interface
Tool granularity	execute_sql()	customer_account_summary() , recommend_films()
Schema knowledge	Discovered from context per request	Implicit in reviewed tool contracts
Joins	Generated per request	Pre-authored and optimized
Business rules	Re-derived by the model	Embedded in server-side SQL
Model tier	Frontier models typical	Small models frequently sufficient
Output shape	Varies per run	Stable, documented columns

Table 1

	Approach	Surface
A	Raw SQL	A single execute_sql tool; the DDL of the six task-relevant tables ( customer , film , category , film_category , inventory , rental ) is embedded in the system prompt. The schema fits entirely in context, so A is not handicapped by prompt size; schema understanding, SQL synthesis and multi-step orchestration are left to the model.
B	Verticalized pack	MCP Blueprint loading packs/sakila v0.5.0: five domain tools ( customer_account_summary , rental_history , recommend_films , film_stock , search_customer ) that accept human-readable identifiers (names, titles) and encapsulate all joins and business rules in pre-authored SQL.
C	Generic thin-tool pack	MCP Blueprint loading a deliberately shallow pack ( search_customer , search_films , get_customer_rentals , get_film ): table-oriented tools with minimal descriptions — a first-pass surface a developer might ship before verticalizing. C isolates the contribution of tool design from tool existence .

Table 2

Model	Params	Tier
llama3.2:3b	3B	SLM
qwen2.5:3b	3B	SLM
qwen2.5:7b	7B	Medium
llama3.1:8b	8B	Medium

Table 3

Category	Tasks
Customer lookup	find_customer , not_found
Rental state & standing	rental_history , return_verify , service_case , overdue_report , good_standing_recommend
Recommendation	recommend_category , recommend_rating , g_available , avoid_on_loan , upsell_seen , not_rented
Catalog details	film_details , store_availability
Multi-step workflow	customer_workflow
Edge cases	rental_empty

Table 4

Metric	A (Raw SQL)	B (Verticalized)	C (Generic)
Mean score	0.666	0.939	0.605
Fully-correct cells	67/201 (33%)	174/204 (85%)	63/204 (31%)
Zero-score cells	7 (3%)	3 (1%)	19 (9%)

Table 5

Model	A	B	C
llama3.2:3b	0.583 (6/51)	0.929 (42/51)	0.419 (0/51)
qwen2.5:3b	0.684 (17/48)	0.902 (42/51)	0.602 (14/51)
qwen2.5:7b	0.647 (20/51)	0.958 (45/51)	0.631 (19/51)
llama3.1:8b	0.750 (24/51)	0.966 (45/51)	0.769 (30/51)

Table 6

Figure 1: Pooled mean accuracy by model and approach. The verticalized pack (B) leads at every model size, with the largest gain on the smallest model (llama3.2:3b), consistent with the Model Demotion hypothesis.

Approach	Mean tokens/cell	vs B	Tokens per correct answer	Seconds per correct answer
B (Verticalized)	3,056	—	3,582	5.2
C (Generic)	2,894	–5.3%	9,372	20.7
A (Raw SQL)	3,953	+29.4%	11,858	51.9

Table 8

Model	A: tokens/correct	B: tokens/correct	A: s/correct	B: s/correct	Reduction
llama3.2:3b	31,476	2,723	49.5 s	3.4 s	$≈$ 11.6 $×$
qwen2.5:3b	17,467	4,766	125.1 s	5.0 s	$≈$ 3.7 $×$
qwen2.5:7b	8,464	4,331	33.2 s	5.6 s	$≈$ 2.0 $×$
llama3.1:8b	5,808	2,531	16.3 s	6.7 s	$≈$ 2.3 $×$

Table 9

Dimension	B (Verticalized)	C (Generic)
Tool semantics	Domain operations ( customer_account_summary , recommend_films )	Table-oriented ( search_films , get_customer_rentals )
Parameter contracts	Human-readable names; descriptions tuned for matching	Bare IDs or minimal labels
Business logic	Encapsulated server-side (standing flag, overdue detection)	Re-derived by the model
Descriptions	Rich guidance steering correct workflows	Minimal, forcing multi-step composition

Table 10

Task	A	B	C	$\Delta$ (B–A)
recommend_category	0.500	1.000	0.833	+50.0pp
avoid_on_loan	0.500	1.000	0.667	+50.0pp
upsell_seen	0.500	1.000	0.750	+50.0pp
film_details	0.562	1.000	0.375	+43.8pp
customer_workflow	0.584	1.000	0.688	+41.6pp
rental_history	0.361	0.694	0.417	+33.4pp

Table 11

Figure 2: Per-task mean score heatmap. Rows are tasks sorted by $\Delta$ (B–A); columns are model $\times$ approach combinations. The verticalized pack achieves near-uniform high scores; raw SQL and the generic pack degrade especially on multi-step workflow and negative-filtering tasks.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

#	Task id	Prompt
1	find_customer	Find the customer whose last name is Smith. Report the customer’s full name and customer ID.
2	rental_history	Mary Smith wants to know what she has rented. What films has she rented, and does she currently have any rentals outstanding?
3	good_standing_recommend	Check whether customer Maria Miller has any overdue rentals. If she is in good standing, recommend two Sci-Fi movies.
4	overdue_report	Check whether customer Tammy Sanders has any overdue rentals. If she does, list the films she still has to return.
5	recommend_category	Recommend three popular Family movies for a family movie night.
6	recommend_rating	Recommend two movies rated G, suitable for all ages.

Appendix

Table 13

LLM Tool Use

Large language model tool use studies how models select, call, and coordinate external tools to extend their reasoning and take actions.

See all 293 papers

MCP Server Architecture Patterns for LLM-Integrated Applications
Jun 29, 2026Carson Rodrigues, Oysturn VasLLM Tool Use Model Context Protocol
Celabe · University of Waterloo
LLM within MCP Matters: Measuring Inefficient Resource Utilization Driven by LLMs
Aug 9, 2026Minhan Cho, Soyoung Park, Kihyeon Jeong +3LLM Inference Efficiency Instruction Following
Remember & Company AI Lab · Sungkyunkwan University · National Assembly Research Service +2
Hybrid Semantic Tool Discovery for Enterprise MCP Gateway: Architecture and Implementation
Aug 25, 2026Olympia Saha, Amy Wang, Srinivasan ManoharanTool Discovery for LLM Agents LLM Tool Use
PayPal, Inc.

Jun 29, 2026J/K move · Enter open · S save

MCP Server Architecture Patterns for LLM-Integrated Applications

Carson Rodrigues, Oysturn Vas

Celabe · University of Waterloo

LLM Tool Use Model Context Protocol

Open paper AskPDF arXiv

The Model Context Protocol (MCP), introduced by Anthropic in November 2024, defines a standardized interface for connecting large language models (LLMs) to external tools, data sources, and services. Within months of release, hundreds of community-built MCP servers appeared on GitHub, but no software-maintenance literature has yet described how the ecosystem is being structured in production. This industry experience paper catalogues five recurring MCP server architectural patterns observed across an enumerated corpus of fifteen independently developed servers (five production servers from the ANSYR voice AI platform plus ten public servers from the official MCP registry): Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter. Each pattern is described in the structured form of Gamma et al.: context, problem, solution, and consequences. We also document four anti-patterns and a set of cross-cutting concerns around authentication, versioning, and observability. The quantitative evaluation contributes three measurements: inter-rater reliability of the taxonomy across two independent LLM raters on 54 held-out servers (Cohen's kappa = 0.76), which also localizes three pattern-boundary ambiguities; transport overhead measured end-to-end on loopback and modeled for cross-host paths; and a tool-count study showing tool-selection accuracy drops below 90% between 10 and 15 tools per context for Claude Haiku 4.5 and between 20 and 30 tools for Sonnet 4. Code, corpus, and prompts are released as a replication package.

Abstract

LLM Tool Use

Explore similar work

MCP Server Architecture Patterns for LLM-Integrated Applications

LLM within MCP Matters: Measuring Inefficient Resource Utilization Driven by LLMs

Hybrid Semantic Tool Discovery for Enterprise MCP Gateway: Architecture and Implementation