promptlibretto 0.5.0

Registry-based prompt model: sections of items, an assembly order, runtime modes, and a browser studio that exports the same JSON the library loads.

promptlibretto

A prompt-engineering library plus a browser studio for designing, tuning, and exporting prompts as portable JSON.

A registry is a flat list of named sections (personas, sentiment, examples, prompt_endings, …); each section holds items with their own fields. An assembly_order of tokens (section / section.field / section[expr]) tells the engine how to weave the selected items into one prompt. Per-array runtime modes (all / index:N / random:K / none), section-level random pickers, optional pre_context headings, and conditional text fragments handle all the small "don't show this if the var is empty" details that otherwise pile up as f-string spaghetti.

• Full docs & walkthrough: sockheadrps.github.io/promptlibretto
• Design rationale: DESIGN.md
• Studio: studio/

Install

pip install promptlibretto                # library only
pip install "promptlibretto[ollama]"      # adds httpx for OllamaProvider
pip install "promptlibretto[studio]"      # adds the browser studio
pip install "promptlibretto[dev]"         # pytest + pytest-asyncio

Two paths

1. Tune in the studio, load JSON in your app

pip install "promptlibretto[studio,ollama]"
promptlibretto-studio --port 8000

Open http://localhost:8000, click Import JSON… to paste a registry (or start fresh), pick a persona / sentiment / examples, set runtime modes (random pick, specific value, or skip with none), fill in any template-var inputs, and click Pre-generate to see the assembled prompt or Generate to send it through your local Ollama.

When you like the setup, click Export Model JSON to copy the canonical registry — including selections, runtime modes, slider positions, and generation overrides. Then in your app:

import asyncio
from promptlibretto import load_registry, OllamaProvider

eng = load_registry(
    "twitch_chatter.json",
    provider=OllamaProvider("http://localhost:11434", "/api/chat"),
)

async def main():
    result = await eng.run(state={
        "selections": {
            "sentiment": "positive",
            "personas": "the_troll",
        },
        "array_modes": {
            "sentiment": {"nudges": "random:1", "examples": "none"},
            "examples":  {"items": "random:3"},
        },
        "section_random": {"personas": True},          # re-roll persona each run
        "sliders":        {"sentiment": 8},            # sentiment.scale token
        "template_vars":  {"base_context::location": "the kitchen"},
    })
    print(result.text)

asyncio.run(main())

load_registry() accepts a path, a JSON string, or a dict. It returns an Engine you can call hydrate() on (no LLM, returns the prompt string) or run() on (hydrate + provider call + output policy).

2. Build it in code

import asyncio
from promptlibretto import Engine, MockProvider, Registry

reg = Registry.from_dict({
    "registry": {
        "assembly_order": ["output_prompt_directions", "personas.context"],
        "output_prompt_directions": {
            "required": True,
            "items": [{"name": "rules", "text": "Be brief."}],
        },
        "personas": {
            "required": True,
            "items": [
                {"id": "shy",   "context": "You're nervous about speaking."},
                {"id": "loud",  "context": "You're full of energy."},
            ],
        },
    }
})

eng = Engine(reg, provider=MockProvider())
print(eng.hydrate(state={"selections": {"personas": "loud"}}))
# Be brief.
#
# You're full of energy.

The full state shape (all fields optional):

state = {
    "selections":     {section_key: id | [id, ...]},   # required: id, multi: [ids]
    "array_modes":    {section_key: {field: "all" | "none" | "index:N" | "random:K"}},
    "section_random": {section_key: bool},             # re-roll item at hydrate time
    "sliders":        {section_key: int},              # drives `<section>.scale` token
    "slider_random":  {section_key: bool},             # re-roll slider at hydrate time
    "template_vars":  {"<section>::<var>": "value"},   # filled into {var} placeholders
}

Optional routes swap assembly_order / generation / output_policy per call:

reg.routes["short"] = Route(assembly_order=["output_prompt_directions"])
result = await eng.run(state=..., route="short")

If a registry has no routes, the top-level assembly_order is the only path.

What's in a registry

{
  "registry": {
    "version": 22,
    "title": "Twitch Chatter",
    "assembly_order": [
      "output_prompt_directions",
      "base_context.text",
      "persona.context",
      "sentiment.context",
      "sentiment.nudges",
      "sentiment.scale",
      "examples.normal_examples",
      "sentiment.examples",
      "prompt_endings"
    ],
    "base_context": {
      "required": true,
      "template_vars": ["location"],
      "items": [{
        "name": "scene",
        "text": "You're watching a streamer at {location}."
      }]
    },
    "personas": { "required": true, "items": [/* {id, context, base_directives[]} */] },
    "sentiment": { "required": true, "items": [/* {id, context, nudges[], examples[]} */] },
    "static_injections":   { "required": false, "items": [/* {name, text} */] },
    "runtime_injections":  { "required": false, "items": [/*
      {id, text, include_sections[], required}
    */] },
    "output_prompt_directions": { "required": true, "items": [/* {name, text} */] },
    "examples":            { "required": false, "items": [/*
      {name, items[], pre_context}
    */] },
    "prompt_endings":      { "required": true, "items": [/* {name, items[]} */] }
  }
}

Sections you don't need can be omitted entirely; the registry is open.

Why bother

• One mental model — sections of items, an assembly order, and per-array runtime modes. No separate route/overlay/injection vocabularies.
• The studio's JSON is the same JSON the library loads. No codegen, no schema drift between editor and runtime.
• Conditional text fragments ({if_var, text}) drop cleanly when their variable is empty, so an unfilled {sublocation} doesn't leave a broken sentence.
• Output policy (length caps, forbidden patterns, required regex, prefix/suffix stripping) and retries on rejection are still here, just nested inside the registry or a route.

Don't use it if you send exactly one prompt shape with one fixed wording. An f-string is fine.

Library API at a glance

Everything below is exported from the top-level package (from promptlibretto import …).

Name	What it is
Engine	Hydrate / run / stream against a registry.
Registry / Section / Route	The model dataclasses. Registry.from_dict / to_dict round-trip the JSON.
HydrateState	Per-call state object. HydrateState.from_dict({...}) accepts the dict shown earlier.
hydrate(reg, state, route, seed)	Functional version of Engine.hydrate. Useful when you don't need a provider.
load_registry(source, provider)	Parse path / JSON string / dict → Engine.
export_json(engine_or_registry)	Serialize back to JSON string.
GenerationResult	text, accepted, prompt, route, reason, usage, timing, raw
GenerationChunk	delta, done, result — yielded by Engine.stream.
GenerationConfig	model, temperature, top_p, top_k, max_tokens, repeat_penalty, retries, max_prompt_chars, timeout_ms.
OutputPolicy / OutputProcessor	Output cleaning + validation rules, see below.
MockProvider / OllamaProvider	Built-in provider adapters.
ProviderAdapter (Protocol)	Implement async def generate(request) -> ProviderResponse to add your own.
StreamingProviderAdapter	Plus def stream(request) -> AsyncIterator[ProviderStreamChunk].
supports_streaming(provider)	True iff the provider has a stream method.
SCHEMA_VERSION / SECTION_KEYS	The current registry version (22) and canonical section keys.

Streaming

async for chunk in eng.stream(state=...):
    if chunk.delta:
        print(chunk.delta, end="", flush=True)
    if chunk.done:
        result: GenerationResult = chunk.result
        print(f"\n[ok={result.accepted} {result.timing}]")

Engine.stream() raises if the provider doesn't implement StreamingProviderAdapter. OllamaProvider does; MockProvider does too.

Custom provider

from promptlibretto import ProviderAdapter, ProviderRequest, ProviderResponse

class MyProvider:
    async def generate(self, request: ProviderRequest) -> ProviderResponse:
        return ProviderResponse(text="…")  # call your API here

eng = Engine(reg, provider=MyProvider())

OllamaProvider configuration

from promptlibretto import OllamaProvider

OllamaProvider(
    base_url="http://localhost:11434",     # default
    chat_path="/api/chat",                 # or "/v1/chat/completions" for OpenAI-compat
    payload_shape="auto",                  # "ollama" | "openai" | "auto"
    timeout=httpx.Timeout(120.0),          # passed to httpx.AsyncClient
)

payload_shape="auto" picks openai when chat_path contains /v1/, ollama otherwise — covers both real Ollama and OpenAI-compatible shims (LM Studio, llama.cpp's server, vLLM).

OutputPolicy fields

All optional, all merged with the registry's top-level policy + the selected route's policy:

OutputPolicy(
    min_length=None, max_length=None,
    strip_prefixes=(),               # remove these prefixes (case-insensitive)
    strip_patterns=(),               # regex; multiline; replaced with ""
    forbidden_substrings=(),         # exact substrings; reject if present
    forbidden_patterns=(),           # regex; reject if matched
    require_patterns=(),             # regex; reject if not matched
    append_suffix=None,              # always appended after cleaning
    collapse_whitespace=True,
)

Sequence-typed fields (strip_prefixes, strip_patterns, forbidden_*, require_patterns) are additive when merged — layered overrides accumulate rules rather than clobbering them. Scalars (max_length, min_length, append_suffix, collapse_whitespace) are replace-on-merge.

Generation overrides at registry / route level

reg = Registry.from_dict({
    "registry": {
        "generation": {"temperature": 0.4, "max_tokens": 96},
        "output_policy": {"max_length": 280, "forbidden_substrings": ['"']},
        "routes": {
            "raid": {
                "assembly_order": [...],
                "generation": {"max_tokens": 64},   # overrides for this route
                "output_policy": {"strip_prefixes": ["Sure,"]},
            },
        },
        # … sections …
    }
})

CLI

promptlibretto-studio --host 0.0.0.0 --port 8000

Environment variables (server-side /api/registry/generate only):

Var	Default	Effect
PROMPT_ENGINE_MOCK	0	1 / true — use MockProvider.
OLLAMA_URL	http://localhost:11434	Base URL for Ollama.
OLLAMA_CHAT_PATH	/api/chat	Chat endpoint path (auto-detects shape).

The studio frontend itself goes browser-direct to the user's local Ollama; these vars only affect the optional server-side generate route.

Development

pip install "promptlibretto[dev]"
pytest

License

MIT (see LICENSE when added).