Domain 3: Understand Copilot Data and Architecture

What this domain is really testing

#

Domain 3 tests the path from local developer activity to an AI-generated suggestion: context collection, prompt assembly, service routing, filtering, model hosting, response handling, and human review.

Think of Copilot as a context-packaging and filtering system around LLMs, not as an editor feature that magically understands your entire repository.

The end-to-end suggestion flow

#

For inline suggestions, Copilot watches the editing context in a supported IDE and offers ghost text or next edit suggestions as you type. The responsible-use documentation describes a pipeline of input processing, language model analysis, response generation, and output formatting; the code suggestions page adds that suggestions matching public code may be discarded or shown with a reference depending on policy. GitHub Docs — Inline suggestions responsible use GitHub Docs — Code suggestions

Developer types or pauses in IDE
        |
        v
IDE extension gathers local context
(current file, cursor area, selected/open relevant files, settings)
        |
        v
Prompt builder ranks and trims context to fit the model context window
        |
        v
GitHub Copilot service / proxy
(authentication, policy checks, content filters, public-code policy)
        |
        v
Model host selected for the feature/model
(OpenAI/GitHub Azure, Anthropic/AWS/GCP, Google Cloud, xAI, Azure, Fireworks, etc.)
        |
        v
Generated candidate suggestion
        |
        v
GitHub Copilot service / proxy post-processing
(harmful/offensive/off-topic filtering, public code matching when enabled, formatting)
        |
        v
IDE displays ghost text, next edit, alternatives, or no suggestion
        |
        v
Human accepts, partially accepts, edits, dismisses, or ignores

The model does not directly read your disk. The IDE extension and Copilot service choose what to send. Selected context becomes input tokens; generated text becomes output tokens; and some models also account for cached tokens. Since 1 June 2026, current Copilot billing is per token under AI Credits, while code completions and next edit suggestions are still not billed in AI Credits. GitHub Docs — Models and pricing

What context Copilot gathers

#

For inline suggestions, Copilot preprocesses surrounding code at the cursor and may add snippets from open tabs. For Chat, documented context can include open files, active repository data, chat history, and optional web search. Repository indexing improves context-enriched answers for Copilot Chat and cloud agent, and indexed repositories are not used for model training. GitHub Docs — Repository indexing GitHub Docs — Prompt engineering

GitHub does not publish an exact numeric scoring formula for context priority. For exam purposes, use this practical order: explicit user-provided context first, then the active editing location, then nearby and open-file context, then broader repository or indexed context, then optional external context such as web search or MCP tools when a feature and policy allow it.

Opening relevant files and closing irrelevant ones is not superstition — it changes the context Copilot is likely to package.

Prompt assembly, tokenization, and context windows

#

The prompt builder packages the request plus selected context into a tokenized model input. Tokens are not words; they are the accounting units used by models and Copilot pricing. The context window limits how much input and working context fits in one interaction, so Copilot must rank, trim, summarize, or omit excess context.

Some current Copilot models support a 1 million token context window in VS Code and Copilot CLI, but that is an extended capability, not whole-repository awareness. Larger context also consumes more tokens and more AI Credits for billable features. GitHub Docs — Supported AI models

Proxy filtering and post-processing

#

The GitHub Copilot service sits between the client and the model host. GitHub’s model-hosting page states that input prompts and output completions continue to pass through GitHub Copilot content filtering systems, including public code matching when applied and mechanisms to detect and block harmful or offensive content. The supported-models page also describes filters for harmful, offensive, or off-topic content and public code matching when enabled. GitHub Docs — Model hosting GitHub Docs — Supported models

Public code matching is setting-dependent. If blocked, Copilot compares suggestions plus about 150 characters of surrounding code against public GitHub code and suppresses matches or near matches in most products. If allowed, or if block mode is unsupported, Copilot can show references with repository URLs and license information. Private repositories and non-GitHub code are not part of that public-code search, and the index refreshes every few months. GitHub Docs — Code referencing GitHub Docs — Finding matching code

Post-processing turns candidate output into ghost text, a next-edit location, a formatted chat response, a code reference, or no suggestion. Nothing is applied until the user accepts it.

Retention, training, and sharing

#

For Copilot Business and Copilot Enterprise, the verified rule to memorize is simple: code snippets are not retained after a suggestion is returned, and customer code is never used to train models. GitHub also documents that it does not use Business or Enterprise customer data to train AI models, while individual Free, Pro, Pro+, and Max interactions may be used for training unless the user disables that setting. GitHub Docs — Model training and improvements GitHub Copilot Trust Center

A fixed Copilot Chat retention period and a transport-security protocol version are outside the verified GH-300 facts used here.

Where models are hosted

#

Copilot can route to models from several providers, but the Copilot service and filters remain in the path. Current OpenAI models listed by GitHub are GPT-5 mini, GPT-5.3-Codex, GPT-5.4, GPT-5.4 mini, GPT-5.4 nano, GPT-5.5, GPT-5.6 Luna, GPT-5.6 Sol, and GPT-5.6 Terra. GitHub also lists Claude, Gemini, Grok, Kimi, Microsoft MAI-Code, and Raptor mini models. GitHub Docs — Model hosting GitHub Docs — Supported models

Auto, base, and LTS models

#

Auto model selection is a routing system that considers system health, availability, task complexity, model policies, subscription type, data-residency restrictions, and FedRAMP-compliance restrictions. It is generally available in Copilot Chat on GitHub.com and VS Code, Copilot CLI, cloud agent, and the GitHub Copilot app. GitHub Docs — Auto model selection

For Copilot Business and Enterprise, a base model is the default model Copilot uses when no other models are enabled. A new base model is automatically enabled within 60 days after designation, giving administrators time to update IDE extensions. An LTS model is supported for one year from designation so organizations can build around a stable target. On 18 March 2026, GitHub designated GPT-5.3-Codex as both the base model and the LTS model. GitHub Docs — Base and LTS models

Limitations you must be honest about

#

Copilot suggestions are probabilistic. They can be plausible and wrong, syntactically valid but semantically broken, biased toward common patterns in training data, stale relative to new libraries or APIs, or inconsistent between runs. The responsible-use pages call out hallucination, inaccurate code, security risks, public-code matches, language-quality differences, and reduced quality when too much context cannot fit into the model call. GitHub Docs — Inline suggestions responsible use GitHub Docs — Chat responsible use

No current Copilot feature should be treated as omniscient whole-repository awareness. Repository indexing and larger context windows improve retrieval, but Copilot still sees a selected slice of context and generates likely output from it. The safe workflow is to review, test, scan, and adapt every important suggestion before it becomes production code.

The exam answer is usually: Copilot can accelerate the work, but the developer remains responsible for correctness, security, licensing, and fit.

Check yourself

#