Request Body Parameters

This page documents every parameter accepted by POST /v1/chat/completions (and the router alias endpoints).

Quick Reference

Parameter	Type	Default	Purpose
model	string	auto	Model preference or routing mode
messages	array	—	OpenAI-style conversation
prompt	string	—	Shortcut when messages is empty
context	boolean\|string	auto	Memory injection mode
chat_id	string	chat_default_<api-key-row-id>	Which memory thread to use
conversation_id / thread_id / memory_id	string	—	Aliases for chat_id
new_chat	boolean	false	Start a fresh random chat
stream	boolean	false	SSE streaming
temperature	number	provider default	Sampling temperature
max_tokens	number	2048 / 4096	Max answer length

---

model

Type: string

Supported values

Value	Behavior
auto / router	Auto-route: classify prompt → pick best provider/model automatically (checks Zen and Nvidia first).
zen / zen-auto	Force Zen provider: pick a random Zen model from the classified tier.
puter / puter-auto	Force Puter provider: skip Zen/NVIDIA, go directly to Puter fallback.
nvidia / nvidia-auto	Force NVIDIA provider: try NVIDIA models only, then fall back to Puter.
<exact-model-name>	Use a specific model (e.g. "deepseek-v4-flash", "gpt-4o"). Routed to Zen first, then Nvidia, then Puter.

Available Zen models

Model	Variants
deepseek-v4-flash-free	low, medium, high, max
mimo-v2.5-free	low, medium, high
big-pickle	(no variants)
laguna-s-2.1-free	low, medium, high
nemotron-3-ultra-free	low, medium, high
north-mini-code-free	none, high

Note

If the model value is omitted or unrecognized, the Worker defaults to auto-routing behavior.

Examples

{ "model": "auto", "messages": [...] }
{ "model": "claude-sonnet-4-6", "messages": [...] }
{ "model": "gpt-4o", "messages": [...] }

---

messages

Type: array

OpenAI-style messages:

[
  { "role": "system", "content": "You are concise." },
  { "role": "user", "content": "Hello" }
]

Supported roles are not strictly validated by the Worker, but the useful roles are:

system
developer
tool
user
assistant

For memory injection, only these are treated as conversation turns:

user
assistant

System, developer, and tool messages are preserved as instruction messages.

---

prompt

Type: string

Shortcut used when messages is missing or empty.

Example:

{
  "model": "router",
  "prompt": "Write a haiku about servers."
}

The Worker converts it into:

{
  "messages": [
    { "role": "user", "content": "Write a haiku about servers." }
  ]
}

---

context

Type: boolean | string Default: auto

Supported values

Value	Behavior
true or "auto"	Store and inject compact memory.
"full"	Store and inject a larger recent-message slice.
false	Clear/skip memory for the selected chat id.

If context is omitted, the Worker sets it to auto.

---

chat_id

Type: string

Use this to continue a specific remembered chat.

If not provided, the Worker uses a stable default chat id for the API key:

chat_default_<api-key-row-id>

---

conversation_id, thread_id, memory_id

Aliases for chat_id.

The Worker checks them in this order:

chat_id
conversation_id
thread_id
memory_id

---

new_chat

Type: boolean

Set to true to create a new random chat id:

{
  "new_chat": true,
  "messages": [
    { "role": "user", "content": "Start a new conversation." }
  ]
}

The response includes the generated chat_id. Save it and send it later to continue the same memory thread.

---

stream

Type: boolean

Set to true for server-sent event streaming.

The Worker supports both upstream formats:

NVIDIA SSE
Puter NDJSON

The client always receives OpenAI-style SSE chunks.

---

temperature

Type: number

Forwarded to the answer provider.

NVIDIA detail

NVIDIA rejects temperature: 0, so the Worker clamps NVIDIA temperatures to at least 0.01. This keeps classifier behavior nearly deterministic while avoiding NVIDIA API validation errors.

---

max_tokens

Type: number

Forwarded to the answer provider.

Defaults

For routed requests, if omitted, the Worker defaults to:

2048 for non-streaming requests
4096 for streaming requests

The classifier itself uses:

max_tokens: 16

Note

If a client sends a smaller max_tokens, that client-provided limit wins. Long answers can still be cut off if you explicitly set max_tokens too low.