RightNow is the model name; System Zero is its category. The full name is RightNow System Zero.
The API has no tool calling: its request schema is strict and rejects tool-calling fields. The SDK provides "RightNow as a tool" for other models through the function-tool helpers and the MCP server adapter (the server ships in the TypeScript package).
Synchronous server-side Python client for the RightNow System Zero API.
Requires Python 3.10 or later.
Install with pip install rightnow-system-zero; import it as system_zero.
Set SYSTEM_ZERO_API_KEY in the server environment, then call the API:
from system_zero import SystemZero
client = SystemZero() # Reuse for the process lifetime; close at shutdown.
reply = client.decide({
"model": "system-zero",
"state": {"order_total": 42},
"questions": {
"eligible": {
"type": "boolean",
"instructions": "Is the order total at least 25?",
},
},
})
print(reply["answers"])
print(client.models.list()["data"])The client calls POST /system-zero/decide and authenticated GET /models.
Boolean, choice and score questions use the same request. decision_for and
decisions_for read typed answers and per-question refusals. Responses preserve
the service's wire fields.
Use decide_with_response or models.list_with_response for response metadata,
including request IDs, rate limits, latency and replay status. Decide options
accept idempotency_key. Both calls accept client_request_id, extra headers
and retry_budget; model listing has no idempotency-key option. Client timeouts
and retry waits are seconds: timeout=150.0, max_retries=5,
retry_budget=60.0 and max_retry_wait=60.0. Close the client or use its context
manager to release the HTTP transport.
Request bodies of 32 KB or more are sent gzip-compressed when that makes them at least 10 percent smaller; pass compress_requests=False to turn this off.
Check models.list() or the documented models endpoint for current input
availability and limits before sending an attachment. SDK types and encoders do
not enable a held API class. Use the public system-zero model ID; an unavailable
explicit model ID returns model_not_found.
Audio is served like every other attachment kind. For any kind, a 503 with
Retry-After is retried after that delay with the same Idempotency-Key;
the SDK does this within its retry limit. If the delay exceeds that limit,
wait for the reported delay, then retry the same request with the same key,
exposed on the error.
attachments is optional; null remains valid. Each inline attachment has a unique
name, a type and a canonical base64 data URL in data. Names start with an
ASCII letter and contain at most 64 letters, digits, _ or -.
Function tools and the MCP adapter carry the inline forms listed by their schema,
including PDF and MP4. They use the same attachment kinds as the published request;
check model discovery for current API availability.
- Images: static PNG (
image/png) and JPEG (image/jpeg), at most 4 images, 16,777,216 pixels each, with an aspect ratio no greater than 50:1. - Audio: one inline WAV (
audio/wav) per request, with integer PCM or IEEE float samples at exactly 16,000 Hz, mono or multichannel. The decoded-duration limit is 60 seconds inclusive (60,000 ms), one pass per request; byte limits can be reached earlier. Audio must be the only attachment. A mixed request is refused locally with HTTP 404audio_not_available, paramattachments. Audio answers currently reportcalibrated: false. Readmodels.list()for current availability and limits. Other audio formats, sample rates and upload references are not supported by this SDK's public audio preflight. - All decoded inline bytes together: at most 2,621,440 bytes. The entire JSON request, including base64, is at most 3,665,920 bytes. The schema has a 16-entry attachment guard; individual kind counts can be stricter.
The helpers check supported MIME types, signatures, canonical encoding, counts, byte limits and WAV structure, sample rate and decoded duration. Preflight is explicit; complete decoding, dimensions, availability and other request checks remain with the API. Signatures alone do not prove a complete file is valid. PDF and video validation remains with the API.
GIF (image/gif), WebP (image/webp), animated PNG and images outside the decoded
profile are not yet supported by this SDK version. The image preflight uses
media_unpriced, media_decode_failed or media_unsupported for unsupported or
mismatched inputs. Upload references are not served: a request naming one answers
HTTP 400 media_unsupported at attachments[i].upload. Send bytes inline.
PDF is supported through ordinary inline attachments.
Text PDFs and readable scans are read as text.
Only PDF is supported by this SDK version; consult discovery for document formats.
Send an unencrypted PDF of at most 500 pages within the common 2,621,440-byte
decoded limit. Other document formats answer HTTP 415 document_unsupported.
Put plain text in state.
Use type: "document" and a canonical data:application/pdf;base64,... URL.
inlineAttachment and inline_attachment can encode a PDF, or construct its
ordinary inline request directly as shown below.
Video is supported through ordinary inline attachments. Send one silent H.264 MP4 video as an inline base64 data URL. Maximum duration: 10 seconds. Maximum dimensions: 1920 x 1080. Maximum decoded file size: 2,621,440 bytes (2.5 MiB). The complete JSON request must fit 3,665,920 bytes, including base64 and questions. Rotated videos, soundtracks, separate attachments and video upload references are not supported.
Use videoAttachment({ name: 'clip', data: bytes }) in TypeScript or video_attachment(name='clip', data=bytes) in Python. Video helpers encode the bytes; the API performs complete-file validation. A video_not_available response becomes VideoNotAvailableError, preserving status, code, message and request ID.
from pathlib import Path
from base64 import b64encode
from system_zero import SystemZero, inline_attachment, validate_attachments
attachments = validate_attachments([
inline_attachment(name="picture", type="image", mime="image/png", data=Path("picture.png").read_bytes()),
])
client = SystemZero() # Create once; reuse across calls and close at process shutdown.
reply = client.decide({
"model": "system-zero", "state": "Use the picture.", "attachments": attachments,
"questions": {"ready": {"type": "boolean", "instructions": "Is the item ready?"}},
})
print(reply["answers"])
# PDF uses the ordinary inline request form.
client.decide({
"model": "system-zero", "state": "Read the PDF.",
"questions": {"ready": {"type": "boolean", "instructions": "Is the item ready?"}},
"attachments": [{"name": "document", "type": "document",
"data": "data:application/pdf;base64," + b64encode(Path("document.pdf").read_bytes()).decode("ascii")}],
})MIMES, MEDIA_LIMITS and attachment_magic_matches expose the SDK encoding registry,
limits and signature screen. inline_attachment encodes one input;
validate_attachments also checks aggregate counts and decoded bytes.
The defaults are five retries, 150 seconds per attempt, a 60-second total retry budget and a 60-second maximum wait. The retry count is capped at five, so one call makes at most six attempts. The recovery budget starts at the first failure and includes retry requests and waits. The first attempt keeps its normal timeout; later attempts are limited by the remaining budget. A wait above the maximum wait, or one that exhausts the remaining budget, raises the last error without waiting. A retry needs at least one millisecond remaining.
Only GET requests or requests carrying a non-empty idempotency key can retry. With retries enabled, decide POSTs receive a generated key when none is supplied. Outside the 1M class, an explicit blank key omits the header and disables decision retries. The 1M class rejects blank keys and generates an omitted key even with retries disabled. Each retry retains the same key and serialized body. Setting the retry count or budget to zero disables retries.
Ordinary HTTP calls retry these marked responses:
- HTTP 429:
capacity_exhausted,rate_limit_exceeded,concurrency_limit_exceededorworkspace_limit_exceeded. - HTTP 5xx except 504 and
maintenance, includingservice_unavailableandsystem_zero_unavailable. - HTTP 409:
idempotency_conflict.
Marked means error.retryable is true or the x-should-retry header is true.
An explicit false in either marker vetoes retry. A cold pool answers marked
503 system_zero_unavailable with its wait in Retry-After.
A qualifying connection failure or timeout before any response byte can also
retry, even if the request was sent. Failures after a response byte never retry.
Errors of authentication and permission types never retry. Python also refuses retry for
codes containing spend, budget, auth or api_key.
Retry-After-Ms takes precedence; only when it is absent does Retry-After
provide seconds or an HTTP date. The wait is never below the full-jitter floor
min(8 seconds, 0.5 seconds * 2^attempt) * random, with attempt zero for the first
retry and random drawn from zero to one. A malformed millisecond header does not
fall back to the seconds header.
run_batch and run_flow use ordinary POST /system-zero/decide calls with a stricter
recovery policy. Function tools use that policy for decide and model
listing. Only a marked HTTP 429 refusal with one of these code/details.cause
pairs proves that no work ran and permits recovery:
rate_limit_exceededwithkey_rateorpreauth.concurrency_limit_exceededwithaccount_concurrency.
The marker is error.retryable=true or x-should-retry: true; an explicit
false in either vetoes recovery. All connection failures and timeouts, HTTP 503
(including warm-up), HTTP 409 and every other code/cause pair are terminal.
The ordinary HTTP retry allowlist above does not expand this workflow policy.
Retries retain one key and one logical call admission and use the client's
retry count, wait and time bounds. Batch retry options can only lower the
client's retry count; they cannot enable retries disabled on the client.
| Helper | Call-cap option | Default | Maximum |
|---|---|---|---|
run_batch |
max_decide_calls |
1,000 | Nonnegative safe integer |
run_flow |
max_decide_calls |
32 | 32 |
create_tool_handler / ToolHandler |
max_decide_calls |
100 | Nonnegative safe integer |
Flow and batch reports record admitted decide calls; retries do not consume
additional logical admissions. Batch receipts also record actual HTTP attempts.
Flows admit a step only when its bounded call fits the remaining deadline;
batches also respect their deadline and the original key's replay lifetime.
Evaluation compares caller-supplied labels locally and adds no network calls.
Explicit call caps must be nonnegative safe integers; zero is allowed. A tool
handler retains its count across calls, so reuse one handler for the tool loop.
Model listing does not consume a decide count. The module-level tool helper has
a zero decide-call limit. A depleted tool limit returns local call_limit_reached
without dispatch; this is separate from API errors.
system_zero.tools supplies function tool definitions and the
ToolHandler, create_tool_handler and handle_tool_call helpers.
They share the API request
validation and error vocabulary. Tool catalog messages are SDK summaries;
HTTP refusal messages are preserved with credential redaction.
SDK errors and response metadata expose the request's idempotency key as
idempotency_key (None when no key was sent). All HTTP errors expose the response's retry delay as retry_after_ms in
milliseconds, or None when no valid retry delay is available.
Every HTTP attempt sends X-SDK-Retry-Count, starting at 0, and a
User-Agent containing only the SDK name, SDK version and runtime version.
A blank or whitespace-only client_request_id omits the
X-Client-Request-Id header.
Keep API keys on trusted servers. Don't enable DEBUG logging for the HTTP or WebSocket libraries in production; it can record request headers, including your API key. See the security policy for private vulnerability reporting.
Create one client per process and reuse it across calls. Do not create a client per call. Close it at process shutdown. Default clients share a process-wide connection pool; closing one client does not close the pool. The transport keeps idle connections for up to 240 seconds, subject to the server closing sooner, so calls after a pause can skip connection setup. The first call after a long pause opens a new connection. An open stream skips connection setup for subsequent decisions; reuse it while sending state changes. A context manager should span the application lifetime.
Breaking: the three function-tool lists were renamed by their JSON shape: NESTED_FUNCTION_TOOLS (nested function objects), FLAT_FUNCTION_TOOLS (flat function objects), INPUT_SCHEMA_TOOLS (input_schema objects). Update imports; the payloads are unchanged.
The three lists describe the same functions. Choose the shape your tool runner accepts:
NESTED_FUNCTION_TOOLS:{"type":"function","function":{"name":"...","description":"...","parameters":{},"strict":true}}.FLAT_FUNCTION_TOOLS:{"type":"function","name":"...","description":"...","parameters":{},"strict":true}.INPUT_SCHEMA_TOOLS:{"name":"...","description":"...","input_schema":{}}.
Each list owns its schema objects. Pass the function name and its JSON arguments to the tool handler.
Supply SYSTEM_ZERO_API_KEY in the server environment before running this example.
The package root and system_zero.stream export SystemZeroStreamClient;
system_zero.stream_interfaces supplies its frame types. The installed runtime
dependency is websockets>=15.0.1,<16. Compression, automatic proxy selection,
and automatic transport pings are disabled; the client manages the gateway's
heartbeat and reconnect contract explicitly.
import asyncio
from system_zero import SystemZeroStreamClient
async def main():
async with SystemZeroStreamClient() as stream:
events = stream.events()
await stream.connect()
await stream.start({
"model": "system-zero", "state": "Order received.",
"questions": {"ready": {"type": "boolean", "instructions": "Is the order ready?"}},
}, decision_mode="on_request", min_interval_ms=100)
await stream.append_text(" Payment received.")
await stream.decide()
async for event in events:
if event["type"] == "decision":
print(event["reply"], event["request_id"])
break
if event["type"] == "error":
print(event["envelope"])
asyncio.run(main())The client sends Authorization: Bearer <key> during the RFC 6455 version 13
handshake and selects only system-zero.stream.v1. Keep the key on a trusted
server. The connection uses wss://api.rightnowai.co/v1/system-zero/stream.
Start with state and questions, then append text, append JSON items, append
attachments, or replace the state. A mutation resolves when its acknowledgement
arrives. In on_change mode, accepted mutations schedule decisions; in
on_request mode, call decide to schedule one explicitly and read its decision
event. One decision can cover several accepted mutations.
Reconnect restores the latest accepted state using stable idempotency keys;
it never resends old wire frames. The default latest policy resumes the selected
decision mode. The never policy restores state in on_request mode, leaving
the next decision to the caller. Neither policy retries authentication or protocol
failures. Idle closure reopens lazily when new work arrives. Listen for
possibly_decided: a lost connection can leave accepted work whose outcome is
unknown, and reopening does not prove that it was uncounted. Opening attempts
are bounded by the retry count and elapsed-time options.
Limits are server-advertised in the hello frame. The SDK paces outgoing state
events using max_events_per_second and the advertised event_burst (or burst)
capacity; when no burst is supplied it conservatively allows one event at a time.
Await each mutation to observe queue backpressure. Waiting for a token does not
start its receipt timeout or block heartbeat replies. The SDK rejects messages
above max_message_bytes before sending. Read hello for session lifetime, idle,
heartbeat, accumulated request and decision interval limits; these can change.
The SDK answers heartbeats automatically. A stream_limit_exceeded refusal
supplies the effective account limit in its details. Respect retry hints.
Billing is per decision, not per connection, heartbeat, acknowledgement, or mutation. Several mutations can be covered by one decision. A decision may be counted even if the connection closes before its event arrives; use the event's usage and request identifier when reconciling completed work.
Tool selection: RightNow System Zero picks which of your tools to call next, or none, with a probability for each. It does not write arguments or messages, and it does not run tools.
Use toolQuestions / tool_questions and selectTool / select_tool. See Tool selection for examples and results.
Text, PDF and video fit 9 tools plus none, or 10 when required. Image or audio fits 25 tools plus none, or 26 when required. One question fits 32,768 bytes and 4,096 tokens. A single required tool is forced locally without a request. Each selection sent is an ordinary answered question on the current prices. Tool selection has not been measured. The caller runs the selected tool and sends its result back in the next request.
Select the class for the state size explicitly. In TypeScript:
await client.decide(request, { dangerZone: 'long_context_64k' });
await client.decide(request, { dangerZone: 'long_context_128k' });
await client.decide(request, { dangerZone: 'long_context_1m' });In Python:
client.decide(request, danger_zone="long_context_64k")
client.decide(request, danger_zone="long_context_128k")
client.decide(request, danger_zone="long_context_1m")long_context_64k covers 38,001 to 65,536 state tokens;
long_context_128k covers 65,537 to 131,072. The option sends
SZ-Danger-Zone with that exact value. State size never selects a class by itself.
Send the text in state with no attachments. A nonempty attachment list with
either class is refused locally as HTTP 404 long_context_not_available.
LongContextRequest excludes attachments. For the 64k and 128k classes, at most 32 questions are allowed,
with 2 to 25 choice options or 2 to 4 score levels. The API checks token bounds.
Use the exact class values above. With accept, text answers HTTP 400
invalid_request, param SZ-Danger-Zone, when more than one native class is served.
With accept, PDF answers HTTP 404 long_context_not_available; image, audio or
video drops the long grants and runs as an ordinary request.
1M needs nothing switched on. It is experimental, not stable. 1M is not guaranteed to be available for every request.
Select long_context_1m explicitly with dangerZone / danger_zone for a
state above 131,072 through 1,048,576 tokens. accept and body size never select
it. Use one string state and exactly one question, with 2 to 25 choice options
or 2 to 4 score levels. Send no attachments, policy, mode, route,
session_id or prompt_cache_key. Empty or null attachments are allowed.
The rendered request is at most 1,048,575 tokens and the body at most 8,388,608
bytes. Use ordinary API-key HTTP calls; this class has no stream or sessions.
Every request carries its full context. Stored answers replay for 24 hours.
The SDK sends your Idempotency-Key or generates one even when retries are
disabled; a blank key is refused locally as HTTP 400 invalid_request, param
Idempotency-Key. Invalid shapes are refused locally as HTTP 404
long_context_not_available; a question count other than one uses
questions_too_large. Nothing needs to be switched on in the Console. A
workspace that has made one completed payment can use the 1M class; the request
option dangerZone / danger_zone (SZ-Danger-Zone: long_context_1m) is still
required. An ineligible workspace gets HTTP 404 long_context_not_available
without a charge; details.eligibility names the first unmet step and may
contain new strings. A state too short for this class is refused with HTTP 404
long_context_not_available, param state, without a charge.
The maximum charge is held before work and settled to exact usage. Insufficient
credit returns HTTP 402 insufficient_credits before work and is not retried
automatically. A workspace that reaches its daily bound for this class gets
HTTP 429 long_context_daily_budget_exhausted with details.scope equal to
"workspace" or "workspace_attempts". The day-wide limit returns the same code
without details.scope. None of these daily refusals is retried automatically;
all clear at the next UTC day. A cold pool returns HTTP 503
service_unavailable before any hold or charge; the SDK retries within its
normal budget using the same key. The first call after an idle period can take several minutes to become available: the answer names the wait in Retry-After; send the same request again with the same Idempotency-Key until it answers. One reply runs at a time per pool; a concurrent
second call returns HTTP 429 capacity_exhausted with Retry-After.
No selector is sent by default. Do not supply the header through custom SDK headers. Retries, batch, flow and their resume paths retain the exact selected class.
Check models.list() and the documented models endpoint for current classes and
workspace requirements. Ordinary limits.state_tokens / rendered_tokens are
the direct window (38,000 / 40,960). Optional limits.long_context lists each
served band with header_value, min_state_tokens, state_tokens,
rendered_tokens and attachments. An ordinary over-length refusal names the
direct window and a selector that covers the state, if available. A selected
class with state outside its band returns non-retryable HTTP 404
long_context_not_available, param state, naming that band; it is not charged.
A held or ineligible class can return HTTP 404
long_context_not_available, raised as
LongContextNotAvailableError. Its eligibility preserves details.eligibility
when supplied, including future rule strings. Held routes are never retried.
Replies type usage.long_input_tokens. Response metadata exposes the server's
sz-danger-zone values as dangerZone in TypeScript and danger_zone in Python.
LongContextLimits describes discovery limits; openapi.json describes the
served contract and enabled.openapi.json contains the additional request shapes.
Every workspace can send policy on ordinary decide calls and stream starts.
min_confidence sets a minimum answer confidence greater than zero and at most
one. Below it, the answer keeps confidence, adds abstained: true and
abstention_reason: "low_confidence", and has no decision member. The reply
echoes policy. An abstained answer is charged exactly like any answered question.
The decision readers return the abstained kind without a decision.
max_cost_usd caps the request's maximum possible cost. If that maximum exceeds
the cap, the API refuses the request with HTTP 400 policy_max_cost_exceeded,
parameter policy.max_cost_usd, without charging it.
on_low_confidence defaults to abstain, its only served value.
on_low_confidence: "error" and require_calibrated are not served: they are
refused with HTTP 400 invalid_request and their respective
policy.on_low_confidence or policy.require_calibrated parameter.
The 1M class does not accept policy.
from system_zero import SystemZero, decisions_for
client = SystemZero() # Reuse this client; close it at shutdown.
reply = client.decide({
"model": "system-zero", "state": "The shipment arrived.",
"questions": {"arrived": {"type": "boolean", "instructions": "Has it arrived?"}},
"policy": {"min_confidence": 0.8},
})
for result in decisions_for(reply).values():
if result.kind == "abstained":
print(result.abstention_reason)
elif result.kind == "v1":
print(result.decision)Read a decision again by its id with the same API key and workspace. Only decides
sent with an Idempotency-Key are kept, for 24 hours. The stored reply is returned
unchanged. Its trace carries no content; some members are null in this version.
An unknown id, an unkeyed decision, another key or workspace, or an expired record
returns HTTP 404 decision_not_found. Reads use the key's request limit, create
no usage and are not charged.
from system_zero import SystemZero
client = SystemZero() # Reuse this client; close it at shutdown.
reply = client.decide({
"model": "system-zero", "state": "The shipment arrived.",
"questions": {"arrived": {"type": "boolean", "instructions": "Has it arrived?"}},
}, idempotency_key="shipment-check")
stored = client.get_decision(reply["id"])
trace = client.get_decision_trace(reply["id"])Every ordinary decide reply carries provenance, including replies delivered by
stream. It adds no cost. The largest context class does not carry it, a reply
too large to carry it omits it, and a decision stored before 2026-11-11 is
returned without it. Check for the member before reading it.
decision_id: use this id for a later decision lookup.request_id: match the original decision request, including after a replay.created: read when the decision was created, as Unix seconds.model: see the model name you requested.resolved_release: see the release that answered.render_id: an opaque identifier of the input rendering.numeric_fingerprint: compare the numeric configuration used for replies.calibration_version: the calibration version for each answered question, as opaque strings.route: see the route that handled the decision.served_by: read the serving identifier returned for the decision.attachment_classes: see the kinds of attachments used, without their content.keyed_replay:trueonly when the reply is a keyed replay (the sameIdempotency-Keysent again);falseon a fresh reply and on a decision lookup.
if reply.get("provenance"):
print(reply["provenance"]["resolved_release"])Create an endpoint for decision.completed, decision.refused or stream.closed.
decision.completed events are sent when the decision's usage is recorded, usually within about five minutes of the answer, and can include decisions answered a few minutes before the endpoint was created.
decision.refused and stream.closed events are sent within about a minute; a new endpoint starts receiving them within 30 seconds.
A decision of the largest context class produces no event in this release. A keyed replay produces no event.
Each delivery carries one signature header: Webhook-Signature: t=<unix seconds>,v1=<64 hex>.
It is HMAC-SHA256 over the ASCII timestamp, a period and the original body bytes.
Verify the raw bytes before parsing JSON. The signature helper accepts timestamps within 300 seconds of the receiver's clock.
Deduplicate deliveries by event id.
The endpoint secret (whsec_...) is shown once when you create the endpoint and once when you rotate its secret. Store it securely then.
Payloads contain no request state, question text, answers' free text or attachments.
An event has an id, type and creation time. Decision data contains decision and request ids, creation time, model, route, input classes and their status, response status, refusal code and usage counters.
Stream-close data contains the stream id, close code and reason.
import json
import os
from system_zero import SystemZero, verify_webhook_signature
with SystemZero(api_key=os.environ["SYSTEM_ZERO_API_KEY"]) as client:
endpoint = client.webhooks.create({
"url": "https://example.com/events", "events": ["decision.completed"],
})
# Store endpoint["secret"] securely now; later reads do not include it.
def handle_delivery(raw_body: bytes, signature_header: str, stored_secret: str):
if not verify_webhook_signature(raw_body, signature_header, stored_secret):
raise ValueError("Invalid signature")
event = json.loads(raw_body)
# Deduplicate by event["id"] before processing event["data"].
return event