Glossary
Plain-language definitions for the AI coding and context-optimization terms that come up most when evaluating coding agents, retrieval methods and token costs.
A
Agentic coding
DefinitionCoding workflows in which an AI system autonomously plans and executes multi-step actions (reading files, running commands, editing code, checking results) toward a goal, rather than only responding to a single prompt with a code suggestion.
Why it mattersThe multi-step, tool-using loop is what makes coding agents token-hungry: every step (search, read, edit, verify) adds to the context sent to the model.
Related termsAI coding agent, vibe coding, context window
AI coding agent
DefinitionA software tool that uses an LLM to autonomously perform software-engineering tasks (writing, editing, testing and debugging code) typically by searching a repository, calling tools, and iterating on results, rather than only producing a single code snippet.
ExampleClaude Code, Cursor and Devin are all AI coding agents.
Related termsagentic coding, MCP, vibe coding
AI coding cost
DefinitionThe total spend incurred running AI coding agents, driven mainly by token consumption (input, output, and any cached or helper tokens) multiplied by a model's per-token price, plus any subscription or usage fees for the agent itself.
Why it mattersCost typically scales with how much context an agent has to read to complete a task, not only with how much code it ends up writing.
Related termstoken, context window, context optimization
CodeMeshCodeMesh's benchmark targets the context and retrieval share of this cost, not the entire AI bill.
AST / Abstract Syntax Tree
DefinitionA tree representation of a program's syntactic structure, produced by parsing source code, where each node represents a construct in the code (a function, a call expression, a loop, an assignment) rather than raw text.
Why it mattersCompilers, linters and code-analysis tools use ASTs because they let tools reason about code structurally instead of as plain text.
Related termsTree-sitter, code graph, structural retrieval
C
Cached token
DefinitionA token served from a model provider's prompt cache rather than freshly processed, typically billed at a lower rate than a normal input token.
ExampleReusing a cached system prompt across requests can cut cost and latency without changing how much context the model actually receives.
Related termstoken, input token, context window
Call graph
DefinitionA directed graph representing which functions or methods call which other functions in a codebase, used to answer questions like "what calls this function" or "what does this function call, transitively."
Related termscode graph, dependency graph, structural retrieval
Code context
DefinitionThe specific code, and code-related information (definitions, call sites, imports, configuration) that an AI coding agent needs in order to complete a given task correctly.
Related termscontext, context optimization, repository context
Code dependency
DefinitionA relationship where one piece of code (a module, function, or package) relies on another to function, whether an internal import within a repository or an external package dependency.
Why it mattersUnderstanding dependencies lets an agent (or a developer) reason about what will be affected by a change before making it.
Related termsdependency graph, call graph, code graph
Code graph
DefinitionA code knowledge graph represents entities in a software repository and the relationships between them. Nodes may represent files, functions, classes or modules, while relationships may represent calls, imports, inheritance, containment or dependencies.
Related termsknowledge graph, call graph, dependency graph, AST / Abstract Syntax Tree
CodeMeshCodeMesh builds and maintains a code graph of a connected repository as the basis for its structural retrieval.
Context
DefinitionIn AI/LLM usage, the information supplied to a model alongside a prompt (prior conversation, retrieved documents, code, or instructions) that the model can draw on when generating a response.
Related termscontext window, code context, context optimization
Context compression
DefinitionReducing the size of the context supplied to a model, through summarization, filtering, deduplication, or retrieving only relevant fragments instead of whole documents or files, while trying to preserve the information needed to complete the task.
Related termscontext optimization, context-window waste, structural retrieval
CodeMeshStructural retrieval is one form of context compression: fetching the specific functions and relationships relevant to a question instead of whole files.
Context engineering
DefinitionThe practice of deliberately designing what information, in what form, and in what order, is supplied to an AI model for a given task (covering prompt structure, retrieved context, tool outputs and conversation history) rather than letting context accumulate in an ad hoc, unmanaged way.
Related termscontext optimization, context, context retrieval
Context optimization
DefinitionContext optimization is the process of minimizing the amount of information supplied to an AI model while preserving the information required to complete a task correctly. For coding agents, this means providing relevant functions, classes, dependencies and code relationships instead of unnecessarily loading large portions of a repository into the context window.
Related termscontext engineering, context-window waste, structural retrieval
CodeMeshCodeMesh is one applied implementation of context optimization for coding agents.
Context pollution
DefinitionThe presence of irrelevant, redundant, or stale information in a model's context that crowds out useful information and can distract or mislead the model, degrading answer quality even when the context window still has room left.
Related termscontext-window waste, context compression
Context retrieval
DefinitionThe process of finding and assembling the specific context (documents, code, prior messages) relevant to a task before it's supplied to a model, whether via search, structural queries, or a retrieval-augmented-generation pipeline.
Related termsstructural retrieval, RAG, semantic search
Context window
DefinitionThe maximum amount of text, measured in tokens, a model can process in a single request, encompassing the prompt, any retrieved context, and the model's own output.
Why it mattersA larger context window allows more information to be supplied, but doesn't by itself make that information relevant or well-organized.
Related termstoken, context-window waste, context
Context-window waste
DefinitionPortions of a used context window occupied by information that doesn't help the model complete the task: duplicate file reads, irrelevant search results, or already-known information re-sent unnecessarily.
Related termscontext pollution, context compression, context optimization
D
Dependency graph
DefinitionA graph in which nodes represent code units (files, modules, packages) and edges represent dependency relationships between them, showing what a piece of code needs to function and what would be affected if it changed.
Related termscall graph, code graph, code dependency
E
Embeddings
DefinitionNumeric vector representations of text or code, positioned in a high-dimensional space so that semantically similar items sit near each other: the basis for vector and semantic search.
Related termsvector search, semantic search, RAG
I
Incremental indexing
DefinitionUpdating an existing repository index by processing only what changed (a saved file, a new commit) instead of reprocessing the entire repository from scratch, keeping an index current without repeated full rebuilds.
Related termsrepository indexing, monorepo
CodeMeshCodeMesh updates its structural understanding of a repository incrementally as files change.
Input token
DefinitionA token in the portion of a model request that represents the prompt and any supplied context, as opposed to the model's generated output, billed by most providers at its own per-token rate.
Related termstoken, output token, cached token
K
Knowledge graph
DefinitionA graph structure representing entities and the relationships between them, used across many domains, not only code, to support structured queries and reasoning that plain keyword search can't. When applied to a codebase specifically, it's often called a code graph.
Related termscode graph, dependency graph
M
MCP
DefinitionModel Context Protocol: an open protocol that lets AI applications (coding agents, assistants) connect to external tools and data sources through a standard interface, so one integration can be reused across multiple compatible clients instead of building a bespoke integration per agent.
Related termsAI coding agent, context retrieval
CodeMeshMCP is the mechanism supported coding agents use to query CodeMesh once a repository is connected.
Monorepo
DefinitionA single repository containing the source code for multiple projects, packages, or services, as opposed to splitting each into its own separate repository.
Why it mattersMonorepos can grow very large, which makes efficient, targeted code retrieval more important for an AI coding agent working inside one.
Related termsrepository context, repository indexing, dependency graph
O
Output token
DefinitionA token in the portion of a model's response that it generates, distinct from the input tokens supplied in the prompt or context, and typically billed at a different, often higher, per-token rate than input tokens.
Related termstoken, input token
R
RAG
DefinitionRetrieval-Augmented Generation: a technique where a system retrieves relevant documents or passages, often via vector or semantic search, and supplies them to a model alongside a prompt, so the model can answer using specific information it wasn't trained on or doesn't reliably recall.
Related termsvector search, embeddings, context retrieval
Repository context
DefinitionThe full set of information available about a codebase that could be relevant to a task (its files, structure, dependencies, history and configuration) as opposed to the narrower slice of that information actually supplied to a model for a specific request.
Related termscode context, context, repository indexing
Repository indexing
DefinitionProcessing a repository's files to build a structured representation of it, such as a code graph or a search index, that can be queried later, instead of re-reading raw files for every request.
Related termsincremental indexing, code graph, monorepo
S
Semantic search
DefinitionSearch that ranks or retrieves results by meaning rather than exact keyword matches, typically implemented using embeddings and vector search so conceptually related text can be found even without shared vocabulary.
Related termsvector search, embeddings, structural retrieval
Structural retrieval
DefinitionStructural code retrieval finds relevant source code using explicit software relationships rather than relying only on textual or semantic similarity. It can retrieve information based on relationships such as function calls, imports, dependencies, inheritance and containment, helping an AI coding agent identify relevant context more directly.
Related termscode graph, call graph, semantic search
CodeMeshThis is how CodeMesh retrieves code: by resolving relationships in its code graph rather than matching text.
T
Token
DefinitionThe basic unit an LLM processes text in, commonly a word, part of a word, or punctuation mark, used both to measure how much text a model can handle in one call (its context window) and, on most platforms, as the basis for usage-based billing.
ExampleA short sentence like "CodeMesh reduces tokens" is typically a handful of tokens, not one per character.
Related termsinput token, output token, cached token, context window
Tree-sitter
DefinitionAn open-source incremental parsing library that builds a concrete syntax tree for source code across many programming languages, used to parse code accurately and efficiently, including as files change, without requiring a full compiler for each language.
Related termsAST / Abstract Syntax Tree, code graph, incremental indexing
CodeMeshCodeMesh uses Tree-sitter to parse supported repositories into its structural representation, across 75 languages: one grammar and one extraction query per language, covering the mainstream set (Python, TypeScript, Go, Java, Rust, C, C++, C#, Ruby, PHP, Kotlin, Swift and the rest) alongside markup, config and IDL formats such as HTML, CSS, SQL, GraphQL, Protobuf, Terraform and Dockerfile.
V
Vector search
DefinitionA search method that finds the nearest items to a query in embedding space, typically using distance metrics like cosine similarity: the mechanism that powers semantic search and RAG systems over unstructured text or code.
Related termsembeddings, semantic search, RAG
Vibe coding
DefinitionA style of software development in which a developer directs an AI coding agent conversationally, describing intent and iterating on results, and relies heavily on the agent to produce and modify most of the code, rather than writing it by hand line by line.
Related termsAI coding agent, agentic coding
See structural retrieval in practice
CodeMesh applies a code graph and structural retrieval to cut the tokens your coding agent spends discovering your repository.