MCP retrieval tools cut tokens 86% on large repos but cost 4x more on small ones

·1 views

A developer ran a weekend experiment comparing MCP retrieval tools against traditional grep-based search across two code repositories of different sizes. On a 33-file TypeScript repository, MCP tools consumed 4.1 times more tokens than grep, while on a 249-file Python and TypeScript repo, they delivered an 86% token reduction. The test used the same AI model and identical tasks across 32 agent runs, with token counts pulled directly from provider responses. The experiment was inspired by the CodeNib research paper, which claimed context-aware retrieval policies could reduce agent trajectory tokens by 50–87%, though the author noted those figures reflect per-model best-case results rather than a single consistent outcome. The findings suggest that MCP-style retrieval tools offer meaningful efficiency gains only beyond a certain repository size threshold.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

SQL vs Cypher: How Graph Queries Simplify Recursive Dependency Lookups

A technical comparison explores how Neo4j's Cypher query language handles complex dependency tree traversals that require lengthy recursive SQL. Using a snapshot of the npm registry's top 5,000 packages, the exercise demonstrates that a 12-line recursive SQL query can be replicated in just two lines of Cypher. The 2016 left-pad incident, where removal of a tiny npm package broke thousands of projects, illustrates why understanding full transitive dependency trees matters. The same dataset was modelled both relationally in SQLite and as a graph in Neo4j, revealing how SQL join tables translate naturally into graph relationships. The article presents ten side-by-side query translations aimed at helping SQL developers read and write Cypher with minimal friction.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Code Really Runs: A Technical Guide to Memory Models and CPU Execution

A detailed technical explainer breaks down what actually happens between writing a line of C code and seeing its output on screen. The piece introduces the memory model as a five-part concept covering data storage, access, lifetime, management, and inter-system communication — going well beyond the common stack-and-heap explanation. It maps the journey of a program across four distinct layers: the programming language, the compiler, the operating system, and the CPU architecture. Using a simple addition program as a running example, it shows how a human-readable variable like 'int x = 10' is progressively translated into assembly instructions that the CPU executes without any concept of named variables. The series aims to give developers a clearer mental model of how memory and execution actually work under the hood.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Trie Data Structures Power Fast Prefix Search and Auto-Complete Features

A Trie is a tree-like data structure where each node represents a character, allowing systems to efficiently handle prefix-based searches such as auto-complete and live search. Unlike HashMaps that retrieve exact values, a Trie organizes data so that words sharing a common beginning also share the same stored path, avoiding duplication. When a user types partial input, the Trie narrows the search space with each additional character rather than scanning the entire dataset. This makes it especially useful in search services, which should be built as dedicated components separate from core business logic. Choosing a Trie over other structures depends on whether the system needs to answer the question: what begins with these characters?

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developers port Python's natsort to Rust, fix 45x slowdown after 133,000 fuzz tests

A development team porting Python's natural-sort library natsort to Rust initially produced a binary nearly four times slower than the interpreted Python original, traced to regex recompilation on every input item. Refactoring the code to compile the regex once per function call cut latency from roughly 900ms to 20ms, making the Rust port 11.8 times faster than Python on identical workloads. A subtle edge case emerged where Python treated bare tokens like 'nAn' as floating-point NaN values during sorting, while the Rust port handled them as plain text, requiring a rewrite of the component-splitting logic. Additional bugs — including a missing NUMAFTER sentinel for numeric-leading strings and an entirely unimplemented PRESORT flag — were uncovered through a custom Pytest bridge that ran the upstream Python test suite directly against the Rust port. After all fixes, the port passed 133,000-plus randomized fuzz inputs with zero mismatches and cleared the Python test suite with no failures.

0 comments Read more at DEV Community