promptlibretto¶

promptlibretto is a prompt registry library. It assembles prompts from named sections, runtime state, and an explicit assembly_order.

Minimal Example¶

from promptlibretto import Engine, MockProvider, Registry

reg = Registry.from_dict({
    "registry": {
        "version": 2,
        "assembly_order": ["output_prompt_directions", "personas.context"],
        "output_prompt_directions": {
            "required": True,
            "items": [{"id": "rules", "text": "Be brief."}],
        },
        "personas": {
            "required": True,
            "items": [{"id": "direct", "context": "Be direct."}],
        },
    }
})

eng = Engine(reg, provider=MockProvider())
print(eng.hydrate({"personas": {"selected": "direct"}}))

Current Public Model¶

• Registry, Section, Route
• RegistryState, SectionState
• ContextItem, Persona, Sentiment, Group
• StaticInjection, RuntimeInjection
• OutputDirection, PromptEnding
• Display, Fragment, Scale

Use these model names when authoring registries in Python.

State Shape¶

State is keyed by section:

{
  "base_context": {
    "selected": "scene",
    "template_vars": {"place": "the kitchen"}
  },
  "personas": {
    "selected": "direct"
  },
  "sentiment": {
    "selected": "warm",
    "slider": 7,
    "array_modes": {"groups[warm_examples]": "random:1"}
  }
}

Assembly Tokens¶

• section
• section.field
• section.scale
• section.groups
• section.groups[group_id]
• groups[group_id]
• injections

Adjacent blocks from the same section join with one newline. Different sections join with two newlines. Adjacent list blocks can merge when they share a pre_context.

Dynamic Items¶

Dynamic items may carry:

• template_vars
• template_defaults
• fragments

Fragments use condition to render only when a section template variable has a value.

Groups¶

Groups are reusable lists of prompt snippets. Attach them to items with groups: ["group_id"], then render with section.groups or section.groups[group_id].

Providers¶

Use Engine.hydrate() for prompt-only rendering, Engine.run() for provider generation, and Engine.stream() for streaming providers.

Built-ins:

• MockProvider
• OllamaProvider

More¶

• Design