osyra_ome.inference

Private pre-release / network boundary: The currently usable customer surface is offline verification of an existing signed artifact against an out-of-band trust anchor. Memory/network RPC APIs and host snippets on this page describe protocol and interface shape; the public Edge bridge, package publication, and release acceptance remain pending. Do not treat api.osyra.ai:443 examples as a released customer route.

Inference-attestation producer over the edge SSE inference stream (ENG-2554 / N6, ADR-284).

The OSYRA inference-binding producer. It drives an inference through the edge's OpenAI-compatible SSE proxy (``POST /api/v1/inference/stream``); when the gateway has OIR emit enabled, the edge mints the Osyra Inference Receipt (OIR) over the attested context + the produced output and returns it as ``oir_b64`` on the terminal SSE frame. This client streams the delta tokens, assembles the output, and surfaces the receipt — the SDK leg of the attested-memory-I/O spine.

Architecture (ADR-285 / the ENG-2546 round-table, Option C): the OIR is minted by the EDGE/OME, NOT the SDK — the SDK sends the request and the gateway (which holds the platform OIR signing key) emits the receipt. The SDK NEVER fabricates a receipt: when emit is dark (the V1 default) ``oir_b64`` is simply ``None`` (NO STUBS). The relying party verifies the receipt OFFLINE under the published ``oirTrustAnchors`` (the frontend ``verifyOirReceipt`` / a future SDK verifier) — never the SDK.

Transport: edge SSE over HTTP (httpx), with the SAME bearer JWT + ``os-workspace-id`` the gRPC ``OmeClient`` / ``GatedMemoryClient`` use; the edge derives the workspace server-authoritatively from the token. Only transport / auth / malformed-response failures raise; a normal stream returns.

Classes

class InferenceClient

A producer client for the edge inference SSE stream + its OIR receipt.

Construct with the edge inference stream URL (``…/api/v1/inference/stream``) + the same bearer JWT the gRPC ``OmeClient`` uses. Pass an ``http_client`` to inject a custom (or mock) transport; otherwise the client owns one and closes it on :meth:`close` / context exit.

Methods

__init__

def __init__(self, inference_url: str, *, token: str | None = None, workspace_id: str | None = None, timeout_seconds: float = DEFAULT_TIMEOUT_SECONDS, http_client: httpx.Client | None = None) -> None

close

def close(self) -> None

Close the owned HTTP client (a no-op for an injected ``http_client``).

complete

def complete(self, model: str, *, prompt: str | None = None, messages: list[InferenceMessage] | None = None, max_tokens: int | None = None, temperature: float | None = None, top_p: float | None = None, stop: list[str] | None = None) -> InferenceResult

Drive an inference to completion and return the assembled output + the OIR receipt.

Consumes the whole SSE stream. Raises :class:`OmeAuthError` / :class:`OmePolicyError` / :class:`OmeTransportError` / :class:`OmeWireFormatError` on auth / policy / transport / malformed failures; a normal stream returns an :class:`InferenceResult` (with ``oir_b64=None`` when the gateway did not emit a receipt).

stream

def stream(self, model: str, *, prompt: str | None = None, messages: list[InferenceMessage] | None = None, max_tokens: int | None = None, temperature: float | None = None, top_p: float | None = None, stop: list[str] | None = None) -> Iterator[str]

Yield each delta token as it arrives (the terminal OIR frame carries no delta).

class InferenceMessage

Bases: BaseModel

One OpenAI-shaped chat turn.

Attributes

| Name | Type | Default | | --- | --- | --- | | content | str | | | role | str | |

class InferenceResult

Bases: BaseModel

The assembled outcome of a streamed inference.

``output`` is the ordered concatenation of the delta tokens (the authoritative output the edge hashes into the OIR — ADR-286). ``oir_b64`` is the base64 OIR receipt when the gateway emitted one, else ``None`` (emit dark / not configured — NEVER a fabricated receipt).

Attributes

| Name | Type | Default | | --- | --- | --- | | finish_reason | str | None | None | | oir_b64 | str | None | Field(default=None) | | output | str | | | usage | TokenUsage | None | None |

class TokenUsage

Bases: BaseModel

Token accounting reported on the terminal frame (may be absent).

Attributes

| Name | Type | Default | | --- | --- | --- | | completion_tokens | int | 0 | | prompt_tokens | int | 0 | | total_tokens | int | 0 |