prism-ai-core

A unified API layer over LLM providers — the Python port of Prism.

Zero runtime dependencies. Python 3.10+.

from prism import Prism

response = (
    Prism.text()
    .using("openai", "gpt-4o")
    .with_prompt("Who are you?")
    .as_text()
)

print(response.text)

Working on this package? Read AGENTS.md first — the boundary this package has to hold, the gates that must be green, and the traps that have already caught someone. @link AGENTS.md

Scope

Stored assistant messages retain provider replay state in additional_content: OpenAI Responses uses output_items; Anthropic uses ordered content_blocks, with thinking_blocks as a fallback. Keep that metadata when sending history back. Signatures, redacted thinking and encrypted reasoning remain opaque and unchanged; avoid logging these values. The legacy thinking/signature pair still works when full replay metadata is absent.

Every capability the reference has, and three of its eighteen providers. That is the shape of the remaining drift: one axis, not two.

• Capabilities — all of them. text, structured, stream, embeddings, images, moderation, audio (text_to_speech and speech_to_text), files, batch, fim. A user message carries images and documents, and each provider spells them its own way.
• Providers — anthropic, mistral, openai. The OpenAI provider speaks the Responses API; Mistral brings the chat-completions shape, which is a different wire format rather than a variant of it.

Not in this port: the tool-execution loop — a response that finishes on tool calls is refused with tool_loop_not_supported rather than half-executed — and the fifteen providers the reference has and this one does not. An unsupported capability raises a coded error rather than a missing attribute.

Not every capability reaches every provider, because not every provider has one. images, moderation, text_to_speech, files and batch are OpenAI-only here, and fim is Mistral-only because Mistral is the only provider that implements it at all — in the reference as well. See the envelope's port gaps register for which of those are work someone could do and which are facts about the provider.

Two things this port has that the reference does not:

• Every failure carries a stable code (prism.ErrorCode). The reference identifies failures by an English sentence, which forces consumers to match on prose. The prose here is explicitly not part of the contract; the code is.
• Every value object rebuilds. to_dict() and from_dict(). The reference can write its value objects and cannot read them back, a gap that forced a downstream package to invent its own rehydration and ship a defect with it.

Install

pip install prism-ai-core

The distribution is prism-ai-core; the package it installs is imported as prism. To track main instead:

pip install "git+https://github.com/Particle-Academy/prism-py"

Do not pip install prism-ai. That name on PyPI is somebody else's package — author gwhite, shipping a prism_ai module, unrelated to this project. It is not a squat to route around later: it publishes the same version number this port carries, so a pin of prism-ai==0.1.0 resolves to the stranger's distribution with nothing to tell you apart.

It is not inert either. It declares httpx, requests and tqdm, where this port declares none, and installing it has been reported ROLLING HTTPX BACK (0.28.1 → 0.25.2) and breaking an MCP client sharing the environment. pip prints that conflict as a warning after the words "Successfully installed", so a CI or otherwise non-interactive install reports success and the damage surfaces somewhere else entirely.

Pin the git URL, and leave a comment where you pin it — this is the kind of line a later tidy-up "corrects" into a version specifier.

Configuration comes from explicit constructor arguments, falling back to OPENAI_API_KEY, OPENAI_URL, OPENAI_ORGANIZATION and OPENAI_PROJECT. The HTTP transport is injectable, so nothing has to reach the network in a test:

from prism import OpenAI, Prism

pending = Prism.text().using("openai", "gpt-4o", {"transport": my_transport})

Parity

What this port must do is pinned by prism-parity, not by this README. The corpus is the contract — the cases, the goldens, the per-language skips and the discrimination probes all live there, each with its own notes saying what it exists to catch. Restating them here would only give them a second copy to drift from.

The corpus and its loader install as the prism-conformance package:

git clone https://github.com/Particle-Academy/prism-parity .parity
pip install ./.parity/loaders/py

Install it last: a path install can be pruned by a later install step. The loader finds its fixtures by walking up from its own installed location, so nothing here ever resolves a path into a sibling checkout.

Run every suite:

python conformance/runner.py

Or one suite, or under a probe:

python conformance/runner.py --suite openai-text-request
python conformance/runner.py --probe omit-null-keys

stdout is JSON and nothing else; the corpus version, digest and root go to stderr on every run. Exit 0 when every case passed or skipped, 1 on a failure, 2 when the corpus failed to load, 3 when the runner could not start.

Probes

conformance/mutations.py implements each probe the corpus declares as an injected defect, and tests/test_probes.py asserts that every one fails exactly the set of case ids the corpus names — not a superset, not "at least one" — and that the faithful control fails nothing. That is what makes the conformance table a measurement rather than decoration.

Nothing in src/prism knows any of this exists. A defect a port can switch on is a defect a port can ship, so the mutants are installed from outside, over the real library's output, and removed again.

Development

src/ layout, so install the package before running the tests:

python -m venv .venv
.venv/Scripts/python -m pip install -e ".[dev]"
.venv/Scripts/python -m pip install ../prism-parity/loaders/py

python -m pytest
python -m mypy --strict
python -m ruff check .
python -m ruff format --check .

Set PRISM_CORPUS_ROOT to run the conformance tests against a parity checkout whose loader copy has not been re-synced yet. Unset — the normal case — the loader discovers its own fixtures.

License

MIT. See LICENSE.