Prompt registry

Version your prompts in one place — with history, rollback without deploy and traceability — and reference them by name. It's opt-in: it coexists with prompts in code (if you don't use it, nothing changes).

from jangada_ai import LLM, PromptVersion

It uses the same config as observability (the project key):

JANGADA_OBSERVABILITY_API_KEY=lobs_xxx
# optional: JANGADA_OBSERVABILITY_ENDPOINT=https://api.jangada.dev.br

Both ways coexist

Prompt in code (default, unchanged):

LLM("openai", "gpt-4o-mini").complete("You are a tax assistant. ...")

Prompt from the registry (opt-in):

p = PromptVersion.pull("tax-assistant")          # production version
LLM("openai", "gpt-4o-mini").complete(p.render(client="ACME"))

In the end the registry resolves to a plain string that goes into the normal complete()/parse() — compatible with {{ }} templates, structured output, tools, etc.

Publish a version (push)

Each push creates a new immutable version. With tag="production", it marks that version as production (moving the tag away from previous ones).

PromptVersion.push(
    "tax-assistant",
    "You are a tax assistant. Answer about {{ topic }} concisely.",
    tag="production",
)

You can also create/version from the dashboard (Prompts tab) — both write to the same registry.

Resolve a version (pull)

pull decides which version to use, in this order:

1. the tag you ask for (PromptVersion.pull("name", tag="staging"));
2. otherwise the one marked production;
3. otherwise the latest version.

p = PromptVersion.pull("tax-assistant")   # production (or latest)
text = p.render(topic="VAT")               # applies the {{ }} template

Rollback

In the dashboard (Prompts tab → the prompt → Make production on an earlier version), the production tag moves back to that version — no deploy. The next pull picks up the restored version.

Trace it ("which prompt generated this?")

Wrap calls in with p.use(): — each trace records which prompt and version generated it (shown in the dashboard, in the call detail, linking to the prompt). Opt-in; requires observability enabled.

p = PromptVersion.pull("tax-assistant")
with p.use():
    resp = llm.complete(p.render(topic="VAT"))   # trace: tax-assistant v2

Why use it

• History + rollback without touching code or deploying.
• Traceability: the team sees which version is in production.
• Comparison via evals: run the same dataset with prompt v2 vs v3 and see which scores higher.
• Autonomy: product tweaks the prompt in the dashboard; the dev isn't a bottleneck.