InstructVault (ivault)

Version prompts in Git, test them in CI, load them locally at runtime.

Prompts live as YAML/JSON files. Changes go through PRs and CI, releases are pinned by tag or SHA, and your app renders them from a local checkout or a bundle artifact — no hosted registry in the request path.

Quickstart

pip install instructvault
ivault init                                              # scaffold prompts/, datasets/, CI workflow
ivault validate prompts                                  # check every prompt spec
ivault render prompts/hello_world.prompt.yml --vars '{"name":"Ava"}'

A prompt looks like this

# prompts/support_reply.prompt.yml
spec_version: "1.0"
name: support_reply
modelParameters:
  model: gpt-4o
  temperature: 0.3
variables:
  required: [ticket_text]
  optional: [customer_name]
messages:
  - role: system
    content: "You are a concise, empathetic support engineer."
  - role: user
    content: |
      Customer: {{ customer_name | default("there") }}
      Ticket: {{ ticket_text }}
tests:                       # at least one test is required
  - name: includes_ticket
    vars: { ticket_text: "My order arrived damaged." }
    assert: { contains_all: ["Ticket:"] }

Use it in your app

render() returns a list of {role, content} messages that also carries the spec's model config, so it drops straight into any client:

from openai import OpenAI
from instructvault import InstructVault

client = OpenAI()
vault = InstructVault(repo_root=".")               # or bundle_path="out/ivault.bundle.json"

result = vault.render(
    "prompts/support_reply.prompt.yml",
    vars={"ticket_text": "My order is delayed", "customer_name": "Ava"},
    ref="prompts/v1.0.0",                          # pin to a tag/SHA (omit for working tree)
)

response = client.chat.completions.create(**result.to_openai())

result is a plain list, so for m in result: m.content still works. Adapters: .to_openai(), .to_anthropic(), .to_litellm(), .to_dict().

CLI

By default eval asserts against the rendered prompt — fully deterministic, no network. Add --provider openai to instead call a model and assert on its reply (needs OPENAI_API_KEY). Network is strictly opt-in, so CI stays deterministic unless you ask for a provider.

Where it fits

flowchart LR
  A[Prompt files] --> B[PR review] --> C[CI validate + eval] --> D[Tag, SHA, or bundle] --> E[App runtime]

Develop locally

git clone https://github.com/05satyam/instruct_vault.git
cd instruct_vault
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
python -m pytest

Docs & examples

• docs/dropin_guide.md — add InstructVault to an existing CI/CD repo
• docs/cookbooks.md — OpenAI, LangChain, LlamaIndex, RAG, policies, bundles
• docs/lockfiles_and_judge.md — reproducible releases + LLM-as-judge evals
• docs/spec.md — prompt spec and validation rules
• docs/ci.md · docs/governance.md · docs/stability.md · docs/roadmap.md · docs/performance.md
• Examples: examples/ivault_demo_template, examples/llamaindex_demo, examples/notebooks, examples/policies

License

Apache-2.0