Builds enriched user prompts with XML-tagged context sections.

Context engineering best practice: place retrieved context (memory, facts, errors) in the user message rather than the system prompt. This keeps the system prompt stable across turns, maximizing KV-cache hit rates with providers that support prompt caching (e.g., Anthropic's prompt caching).

Each context section is wrapped in an XML tag and placed before the user message so the LLM treats it as retrieved context, not user input.

Usage

alias Jido.AI.PromptBuilder

prompt = PromptBuilder.build("What meetings do I have today?", [
  {:memory_context, memory_block},
  {:known_facts, facts_string},
  {:previous_errors, error_string}
])

Sections with nil or empty string values are automatically skipped.

Custom Tags

You can use any atom as a tag name:

PromptBuilder.build("Hello", [
  {:retrieved_documents, docs_text},
  {:user_preferences, prefs_text}
])

This produces:

<retrieved_documents>
...docs...
</retrieved_documents>

<user_preferences>
...prefs...
</user_preferences>

Hello