cyberian-client 0.4.0

Python client for the Cyberian Systems verified-inference API

cyberian-client

Python client for the Cyberian Systems verified-inference API.

Each call returns embeddings + a cryptographic receipt attesting that the inference was correctly executed and independently verified. The receipt is self-contained: any third party can validate it without access to the platform.

Install

pip install cyberian-client

Requires Python 3.9+. Pulls in httpx and numpy.

Get an API key

https://cyberiansystems.ai/signup — 14-day free trial, no card required. The plaintext key is shown once at signup; save it. We store only its SHA-256.

Quickstart — one shot

from cyberian import Cyberian

client = Cyberian(api_key="cyb_trial_…")  # or set CYBERIAN_API_KEY in env

result = client.submit_and_wait(
    model="bge-large-v1.5",
    inputs=[
        "The property title search revealed no encumbrances.",
        "Quarterly compliance attestation per SOX 404.",
    ],
)

print("receipt_hash:", result.receipt["receipt_hash"])
print("verified:    ", result.verification["valid"])  # True
print("shape:       ", result.embeddings.shape)        # (2, 1024) for BGE-large
print("first vec:   ", result.embeddings[0, :8])

Step-by-step — when you need finer control

job = client.submit_job(model="bge-large-v1.5", inputs=["a", "b", "c"])

# Poll yourself, or use wait_for_completion:
final = client.wait_for_completion(job["id"], timeout_sec=600)

receipt      = client.get_receipt(final["receipt_id"])
verification = client.verify_receipt(final["receipt_id"])
embeddings   = client.get_job_output(final["id"])  # numpy ndarray, float32

Models

Every job names a model (model="..."). The models enabled for your account are listed on your account dashboard, or programmatically:

for m in client.list_models():
    print(m["id"], "—", m["name"])

list_models() returns a list of descriptors; the exact fields are stable across releases but may grow over time, so dereference by name rather than position.

Need a model that isn't enabled for you? Email [e-mail] with the model reference (e.g. a Hugging Face URL). Onboarding a new model is a configuration step on our side — no code change on yours.

Verification as a Service (VaaS)

You ran inference on your own infrastructure. Cyberian independently verifies the output you computed and issues a receipt certifying it. Useful for compliance — you keep the inference, we provide independent auditable attestation.

import hashlib

# Your in-house inference output:
my_embeddings_bytes = my_pipeline.run(["text"]).astype("float32").tobytes()
my_commitment = hashlib.sha256(my_embeddings_bytes).hexdigest()

receipt = client.verify_outputs(
    model="bge-large-v1.5",
    inputs=["text"],
    claimed_output_commitment=my_commitment,
)
# Raises VerificationFailedError if our independent verification disagrees with
# the commitment you claimed.

Verifying more than one batch in a single call

For larger workloads, split your outputs into batches of up to 1000 inputs and pass one commitment per batch (in input order) via claimed_output_commitments. You get a single receipt covering every batch:

batches = [inputs[i:i + 1000] for i in range(0, len(inputs), 1000)]
commitments = [
    hashlib.sha256(
        my_pipeline.run(batch).astype("float32").tobytes()
    ).hexdigest()
    for batch in batches
]

receipt = client.verify_outputs(
    model="bge-large-v1.5",
    inputs=inputs,
    claimed_output_commitments=commitments,
)

Pass exactly one of claimed_output_commitment (single batch) or claimed_output_commitments (multi-batch). If the number of commitments doesn't match the batch count, the platform returns a commitment_count_mismatch error telling you how many it expected. Per-request limits depend on your account.

Receipt fields

The receipt your jobs / VaaS calls return is a self-contained, hash-chained object:

Field	Meaning
receipt_id	UUID of this receipt.
job_id	UUID of the underlying job.
spec_hash	SHA-256 commitment to the execution spec used.
inputs_root	Merkle root over the inputs you submitted.
outputs_root	Merkle root over the outputs that were produced.
verification_attestation	Merkle root committing to the independent verification work.
executor_attestation_ids	Anonymized identifiers of the providers that executed the batch.
issued_at	UTC ISO-8601 timestamp.
receipt_hash	SHA-256 of the entire receipt body — self-referential check.

Account management

info = client.get_account()
print(info["tier"], info["effective_tier"], info["subscription_status"])
print(info["chunks_consumed_period"], "/", info["limits"]["chunks_per_period"])

# Mint an additional key (old one stays valid until you revoke it)
new_key = client.rotate_key()

# Revoke a specific key by its 12-char key_id (the prefix shown in /account)
client.revoke_key(key_id="a1b2c3d4e5f6")

Error handling

Typed exceptions; catch the most specific one you care about.

from cyberian import (
    Cyberian,
    AuthError,                # 401, 403
    TrialExpiredError,        # 402 — email [e-mail]
    QuotaError,               # 402 / 413
    RateLimitError,           # 429 — has .retry_after_sec
    ServiceBusyError,         # 503 — has .retry_after_sec
    JobFailedError,           # job ended in FAILED state during wait_for_completion
    VerificationFailedError,  # /verify returned valid=False
    NetworkError,             # transport-level (no HTTP response); wait_for_completion retries automatically
    ApiError,                 # any other 4xx/5xx
    CyberianError,            # base of everything above
)

try:
    result = client.submit_and_wait(model="bge-large-v1.5", inputs=[…])
except RateLimitError as exc:
    time.sleep(exc.retry_after_sec)
    result = client.submit_and_wait(model="bge-large-v1.5", inputs=[…])
except TrialExpiredError:
    print("Email [e-mail] for an extension or enterprise tier.")
    raise

Trial limits

	trial (default)	free (post-trial)
Period	14 days	30 days
Chunks per period	400	100
Chunks per day	100	25
Requests per minute	60	10
Max chunks per request	100	50
Max inputs per request	1000	500

A "chunk" is the platform's internal billing unit (one batched inference). The platform chooses how to group your inputs into chunks. Counting chunks (not raw requests) means the quota tracks actual compute consumption.

Documentation

• API + SDK reference: https://cyberiansystems.ai/docs
• Trial terms: https://cyberiansystems.ai/terms
• Privacy: https://cyberiansystems.ai/privacy
• Issues / support: [e-mail]
• Upgrade / enterprise inquiries: [e-mail]