galene-ai

Python SDK for the Galene.AI platform — a single, typed client covering all 317 REST operations of the API, generated and hand-crafted from the vendored OpenAPI spec (spec/openapi.json).

• Familiar ergonomics — client.chat.create(...), client.files.upload(...), OpenAI-SDK-style.
• Sync and async — Galene and AsyncGalene share one core and expose an identical API.
• Typed — msgspec models generated from the spec; mypy --strict clean.
• Slim — two runtime dependencies: httpx and msgspec.
• Production-ready — retries with backoff, pluggable auth (API key or username/password with auto-refresh), SSE streaming, cursor pagination.

Install

pip install galene-ai

Requires Python 3.12+.

Quickstart

from galene_ai import Galene
from galene_ai.models._generated import VectorStoreCreate, VectorStoreFileAdd, VectorStoreSearchRequest

client = Galene(api_key="sk-...")  # or set GALENE_AI_API_KEY in the environment

# Chat completion
reply = client.chat.create(
    model="Galene/LLM",
    messages=[{"role": "user", "content": "Hi"}],
)
print(reply.choices[0].message.content)

# Upload a file
uploaded = client.files.upload(b"some file bytes", purpose="user_data", filename="doc.txt")

# Vector stores (RAG)
vector_store = client.vector_stores.create(VectorStoreCreate(name="kb"))
client.vector_stores.add_file(vector_store.id, VectorStoreFileAdd(file_id=uploaded.id))
results = client.vector_stores.search(vector_store.id, VectorStoreSearchRequest(query="pricing"))

client.close()

VectorStoreCreate, VectorStoreFileAdd, and VectorStoreSearchRequest are generated request models (see examples/vector_store_rag.py for the full, runnable version). Galene is also a context manager:

with Galene(api_key="sk-...") as client:
    ...  # client.close() is called automatically

Authentication

The default is a static API key, supplied explicitly or via environment variable:

from galene_ai import Galene

client = Galene(api_key="sk-...")   # explicit
client = Galene()                   # reads GALENE_AI_API_KEY from the environment

Environment variables:

Variable	Purpose	Default
GALENE_AI_API_KEY	API key used by the default ApiKeyAuth	—
GALENE_AI_BASE_URL	API base URL	https://api.playground.galene.ai

Both are optional — the base URL defaults to the public playground, so point it at your own deployment via the env var or the base_url= argument. The SDK also auto-loads a .env file from the working directory (real environment variables always win). Copy .env.example to .env and fill it in:

cp .env.example .env   # then set GALENE_AI_API_KEY / GALENE_AI_BASE_URL

For username/password login with transparent token refresh, pass a SessionAuth instance as auth=:

from galene_ai import Galene, SessionAuth

client = Galene(auth=SessionAuth(username="me@example.com", password="hunter2"))

SessionAuth logs in via POST /login on the first request, attaches the resulting access token to every request, and — on a 401 — transparently calls POST /refresh-token and retries once. Auth is pluggable: any httpx.Auth subclass can be passed as auth=.

Async usage

AsyncGalene mirrors Galene method-for-method:

import asyncio
from galene_ai import AsyncGalene


async def main() -> None:
    async with AsyncGalene(api_key="sk-...") as client:
        reply = await client.chat.create(
            model="Galene/LLM",
            messages=[{"role": "user", "content": "Hi"}],
        )
        print(reply.choices[0].message.content)


asyncio.run(main())

Streaming

Pass stream=True to chat.create (or responses.create) to get a Stream[dict] (or AsyncStream[dict]) of raw SSE chunks — the spec defines no fixed chunk schema, so chunks decode as plain dicts in the standard OpenAI delta shape:

with client.chat.create(
    model="Galene/LLM",
    messages=[{"role": "user", "content": "Count to five."}],
    stream=True,
) as stream:
    for chunk in stream:
        delta = (chunk.get("choices") or [{}])[0].get("delta") or {}
        print(delta.get("content") or "", end="", flush=True)

The async equivalent uses async for chunk in stream: inside async with.

Pagination

List methods on cursor-paginated namespaces (files, vector_stores, ...) return a CursorPage[T] (or AsyncCursorPage[T]) with .data, .has_more, and .auto_paging_iter():

for vector_store in client.vector_stores.list().auto_paging_iter():
    print(vector_store.id, vector_store.name)

Async (async for on an AsyncCursorPage already walks every page — no separate auto_paging_iter() call needed):

async for vector_store in await client.vector_stores.list():
    print(vector_store.id)

Namespaces that use offset pagination (e.g. observability) instead expose plain limit/offset keyword arguments and return a list[...] directly.

Error handling

Every error raised by the SDK subclasses GaleneError. HTTP error responses raise a status-specific subclass of APIStatusError, mirroring the OpenAI SDK hierarchy:

from galene_ai import GaleneError, NotFoundError

try:
    client.files.retrieve("file-does-not-exist")
except NotFoundError as e:
    print(e.status_code, e.request_id, e.body)
except GaleneError as e:
    print("Galene API error:", e)

GaleneError
├─ APIConnectionError            (network failure)
│  └─ APITimeoutError
└─ APIStatusError                (.status_code, .request_id, .response, .body)
   ├─ BadRequestError            400
   ├─ AuthenticationError        401
   ├─ PermissionDeniedError      403
   ├─ NotFoundError              404
   ├─ ConflictError              409
   ├─ UnprocessableEntityError   422
   ├─ RateLimitError             429
   └─ InternalServerError        5xx

Connection failures and 408/409/429/500/502/503/504 responses are retried automatically with exponential backoff (max_retries=2 by default, configurable via Galene(max_retries=...)); non-idempotent generation requests are never retried on a read timeout, since a timed-out generation keeps running server-side.

Namespaces

Every operation is reachable as client.<namespace>.<method>(...) (async: AsyncGalene, identical shape). Admin operations nest under client.admin.<namespace>.<method>(...).

OpenAI-compatible surface (30 ops):

Namespace	Ops	Namespace	Ops
chat	2	files	6
responses	2	vector_stores	11
embeddings	2	audio	4
models	2	moderations	1

Platform surface (212 ops):

Namespace	Ops	Namespace	Ops	Namespace	Ops
agents	46	attachments	10	api_keys	6
observability	16	conversations	10	groups	6
tickets	14	shield	9	notifications	5
users	13	mcp_servers	8	roles	4
auth	12			health	2
kb_sync	12			organizations	2
tts	12			release_notes	2
database_connectors	11			changelog	1
kb_connectors	11

Admin surface, nested under client.admin.* (74 ops):

Namespace	Ops	Namespace	Ops
admin.org_config	33	admin.tickets	7
admin.global_config	25	admin.management	5
		admin.deployment	1
		admin.frontend	1
		admin.search	1
		admin.user_api_keys	1

For example:

client.agents.list()
client.observability.list_traces(status="running")
client.admin.org_config.set_favicon(org_uuid, favicon_png_bytes)

observability exposes trace/observation listing, read-audit trails, CSV exports, and aggregate metrics directly (no further sub-resource nesting); conversations exposes message- and attachment-related helpers the same way. Endpoints with mode-prefixed variants (chat, responses, embeddings, models; mode ∈ direct/shield/appliance_shield) expose a create_with_mode(mode, ...) (or list_with_mode(mode)) sibling method alongside the plain /v1/... one.

The full, machine-readable operation → namespace/method mapping lives in spec/operations.json.

Typing note: a handful of endpoints return dict

spec/openapi.json leaves the 200 response schema empty ({}, no $ref) for a number of operations — most notably the entire agents namespace (46 ops) and attachments namespace (10 ops), plus a few others scattered across the SDK (models.list, audio.voices, the observability .csv export endpoints, some vector_stores delete/batch responses). datamodel-code-generator has nothing to generate a model from in those cases, so the SDK decodes the response as a plain dict[str, Any] rather than guessing a shape that might not match production.

To upgrade one of these to a typed msgspec.Struct:

1. Run scripts/probe_shapes.py against a reachable Galene instance (GALENE_AI_API_KEY=... python scripts/probe_shapes.py). It only issues GET requests — it never mutates data — and writes the observed JSON shapes to probe_output/<operationId>.json, plus a summary table to stdout.
2. Add a hand-written msgspec.Struct for the observed shape to src/galene_ai/models/_extra.py (see that module's docstring for the existing pattern used for File, VectorStore, etc.).
3. Update the resource method's cast_to= argument to the new model.

Development

uv sync --extra dev

• Regenerate models (dev-only, output is committed): uv run python scripts/generate_models.py
• Regenerate the operation registry (dev-only, output is committed, drives the drift test): uv run python scripts/build_registry.py
• Tests: uv run pytest
• Type-check: uv run mypy
• Lint / format: uv run ruff check . / uv run ruff format .

tests/test_drift.py asserts that the set of operationIds implemented by resource methods equals the set in spec/operations.json exactly — any API addition, removal, or rename that isn't reflected in the SDK fails CI.

License

MIT