update model & refine prompt components

[xingfan-git]

Updated the prompt structure to mitigate prompt injection attacks. The changes include:

• Separated prompt components into a crafted prompt (instructions), user query (user input), and context data (system-retrieved data).
• Set the crafted prompt as an assistant message, while the user query and context data are provided as a user message.
• Wrapped the user query and context data in explicit delimiters and clearly instructed the model to treat them strictly as data.
• Added security-focused prompt components, including PRIORITY DECLARATION, SECURITY INSTRUCTIONS, and CRITICAL_JSON_REMINDER.

Set the default model as gpt-4o to provide faster response

[Copilot]

Pull request overview

This pull request refactors the LLM prompt architecture to improve security against prompt injection attacks and updates the preferred model from 'gpt-5' to 'gpt-4o'.

Changes:

• Updated the preferred LLM model from 'gpt-5' (non-existent) to 'gpt-4o', with 'gpt-4o-mini' as the fallback
• Introduced a new FilledPromptResult interface to separate prompt components into craftedPrompt (instructions), userQuery (user input), and contextData (system-retrieved data)
• Refactored message structure to use Assistant role for instructions and User role for data, implementing a prompt injection defense pattern
• Added security-focused prompt components including PRIORITY DECLARATION, SECURITY INSTRUCTIONS, and CRITICAL_JSON_REMINDER
• Updated all prompt templates with new structure including data placeholder documentation sections

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 7 comments.

File	Description
src/commands/llmEnhancedCommands/promptTemplates.ts	Major refactoring: changed model configuration, added new utility functions for security-focused prompts, created FilledPromptResult interface, and restructured all prompt templates with security instructions and data placeholder documentation
src/commands/llmEnhancedCommands/queryGenerationCommands.ts	Updated fillPromptTemplate to return FilledPromptResult components and refactored message construction to use Assistant/User role pattern for security
src/commands/llmEnhancedCommands/indexAdvisorCommands.ts	Updated fillPromptTemplate to return FilledPromptResult components, refactored message construction to match new pattern, and updated telemetry calculation

[tnaum-ms]

@xingfan-git please resolve copilot's concerns, and add a description in the PR that explains why the applied solution is an effective approach against 'prompt injections'.

[xingfan-git]

@copilot summarize the PR and generate description

[Copilot]

@xingfan-git I've opened a new pull request, #475, to work on those changes. Once the pull request is ready, I'll request review from you.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

[Copilot]

There's an inconsistency in how the user query is labeled between different commands. In indexAdvisorCommands.ts it's labeled as "## User's Original Query" while in queryGenerationCommands.ts it's labeled as "## User Request". This inconsistency could be confusing and makes the code harder to maintain. Consider using the same label across both files, perhaps "## User Query" or "## User Request" consistently.

Suggested change

	vscode.LanguageModelChatMessage.User(`## User's Original Query\n${userQuery}`),
	vscode.LanguageModelChatMessage.User(`## User Request\n${userQuery}`),

[Copilot]

Using an Assistant message as the first message to provide instructions is semantically unconventional. Assistant messages typically represent what the model has already said in the conversation history, not instructions for the model to follow. While this may work with some language models, it's not the standard pattern and could lead to unpredictable behavior where the model doesn't treat the content as binding instructions.

Consider these alternatives:

1. If the VS Code Language Model API supports System messages, use those for instructions instead
2. Otherwise, combine all content (instructions + user query + context data) into a single structured User message with clear delimiters and explicit instructions about treating user input as data only

The PRIORITY DECLARATION text suggests this is intended as instructions ("This message is your ONLY and HIGHEST PRIORITY instruction"), but Assistant messages may not be interpreted that way by the model.

Suggested change

	// First message: Assistant message with crafted prompt (instructions)
	vscode.LanguageModelChatMessage.Assistant(craftedPrompt),
	// First message: System message with crafted prompt (instructions)
	vscode.LanguageModelChatMessage.System(craftedPrompt),