This page summarizes common request parameters for Moxus AI. Whether a parameter takes effect depends on the protocol, selected model, and upstream provider. Unknown or unsupported fields may be ignored by the provider or rejected with a parameter error.
Do not mix request bodies across protocols. OpenAI-compatible APIs use https://moxus.ai/v1, Anthropic-compatible APIs use https://moxus.ai, and Google-compatible APIs use https://moxus.ai/v1beta.
Common fields
| Parameter | Type | Required | Default | Range / example | Meaning |
|---|
model | string | Yes | None | gpt-4o, claude-3-5-sonnet, gemini-2.5-pro | Model name. Copy the exact name from Model Square. |
messages | array | Required for chat / Claude | None | [{"role":"user","content":"..."}] | Conversation messages. OpenAI and Anthropic message shapes differ. |
input | string / array / object | Common for Responses / embeddings | None | Text, arrays, multimodal input | Main input field for Responses, embeddings, and some models. |
prompt | string / array | Some completion or image APIs | None | "Draw a cat" | Single prompt or image generation prompt. |
stream | boolean | No | false | true / false | Return streaming responses. |
metadata | object | No | None | {"trace_id":"abc"} | Custom metadata for business tracing. Do not put secrets here. |
user | string / object | No | None | "user_123" | Client-side user identifier. Some providers support it; others ignore it. |
Sampling and length
| Parameter | Type | Default | Range / example | Meaning |
|---|
temperature | number | Model default | Usually 0 to 2 | Controls randomness. Lower is more stable; higher is more diverse. |
top_p | number | Model default | Usually 0 to 1 | Nucleus sampling threshold. Avoid changing both temperature and top_p aggressively. |
top_k | integer / number | Model default | Positive number, provider-specific | Limits candidate tokens. Common in Anthropic and Google; not always supported by OpenAI models. |
max_tokens | integer | Model default | Non-negative integer within model/platform limits | Maximum output tokens for Chat Completions and Anthropic-style calls. |
max_completion_tokens | integer | Model default | Non-negative integer within model/platform limits | Output cap for some newer OpenAI models. |
max_output_tokens | integer | Model default | Non-negative integer within model/platform limits | Output cap for Responses API and Google-native calls. |
stop | string / array | None | "\n" or ["END"] | OpenAI-compatible stop sequence. |
stop_sequences | array | None | ["END"] | Anthropic / Google native stop sequences. |
n | integer | 1 | Usually positive integer | Number of candidates. More candidates usually cost more. |
seed | number / integer | None | 0 or another integer | Attempts deterministic output. Reproducibility depends on model and provider. |
OpenAI Chat Completions parameters
| Parameter | Type | Default | Range / example | Meaning |
|---|
messages | array | None | system / user / assistant / tool | Conversation messages. |
response_format | object | None | {"type":"json_object"}, {"type":"json_schema",...} | Ask the model to return JSON or schema-constrained output. |
tools | array | None | [{"type":"function","function":{...}}] | Declares tools the model may call. |
tool_choice | string / object | auto | auto, none, or a specific function | Controls whether or which tool is called. |
parallel_tool_calls | boolean | Model default | true / false | Allows multiple tool calls in one response. |
logprobs | boolean | false | true / false | Returns token log probabilities if supported. |
top_logprobs | integer | None | Usually 0 to 20 | Number of candidate token logprobs per position. |
frequency_penalty | number | 0 | Usually -2 to 2 | Penalizes repeated token frequency. |
presence_penalty | number | 0 | Usually -2 to 2 | Encourages new topics when positive. |
logit_bias | object | None | { "123": -100 } | Adjusts likelihood for token IDs. |
modalities | array | Model default | ["text"], ["text","audio"] | Output modalities, depending on model support. |
audio | object | None | {"voice":"alloy","format":"mp3"} | Audio output configuration. |
OpenAI Responses parameters
| Parameter | Type | Default | Range / example | Meaning |
|---|
input | string / array | None | Text or multimodal input array | Main Responses API input. |
instructions | string | None | "You are helpful" | Global instructions, similar to a system prompt. |
previous_response_id | string | None | resp_... | Continue from a previous Responses call. |
max_output_tokens | integer | Model default | Non-negative integer | Maximum output tokens. |
reasoning | object | Model default | {"effort":"medium"} | Reasoning configuration for reasoning models. |
text | object | Model default | {"format":{...}} | Text output format configuration. |
tools | array / object | None | Web search, function, MCP, and more | Responses tool configuration. |
tool_choice | string / object | auto | auto, none, or a specific tool | Controls tool use. |
truncation | string | Model default | auto, disabled | Behavior when input is too long. |
include | array | None | [...] | Requests additional response fields. |
store | boolean | Upstream default | true / false | Whether upstream may store output. Channel settings may control passthrough. |
Anthropic-compatible parameters
| Parameter | Type | Required | Default | Range / example | Meaning |
|---|
model | string | Yes | None | claude-3-5-sonnet | Claude or Anthropic-compatible model name. |
messages | array | Yes | None | user / assistant messages | Claude Messages content. |
system | string / array | No | None | "You are helpful" | System instruction. |
max_tokens | integer | Commonly required | None | Non-negative integer | Maximum output tokens. |
temperature | number | Model default | Usually 0 to 1 | Controls randomness. Some Claude models may reject non-default sampling parameters. | |
top_p | number | Model default | Usually 0 to 1 | Nucleus sampling. | |
top_k | integer | Model default | Positive integer | Candidate token limit. | |
stop_sequences | array | None | ["END"] | Stop sequences. | |
stream | boolean | false | true / false | Return streaming response. | |
tools | array | None | Tool declarations | Declares Claude tools. | |
tool_choice | object | auto | {"type":"auto"}, {"type":"tool","name":"..."} | Controls tool selection. | |
thinking | object | Model default | {"type":"enabled","budget_tokens":1024} | Claude thinking configuration, depending on model support. | |
metadata | object | None | {"user_id":"..."} | Business metadata. Avoid sensitive data. | |
Google-compatible parameters
Google-native APIs usually put generation parameters in generationConfig. Moxus AI also accepts some snake_case fields, such as top_p and max_output_tokens.
| Parameter | Type | Default | Range / example | Meaning |
|---|
contents | array / string | None | Text or multimodal parts | Main Google-native input. |
systemInstruction | object / string | None | System instruction | Google-native system instruction. |
tools | array / object | None | functionDeclarations, googleSearch, etc. | Tool or search declarations. |
generationConfig.temperature | number | Model default | Usually 0 to 2 | Controls randomness. |
generationConfig.topP / top_p | number | Model default | Usually 0 to 1 | Nucleus sampling. |
generationConfig.topK / top_k | number | Model default | Positive number | Candidate token limit. |
generationConfig.maxOutputTokens / max_output_tokens | integer | Model default | Non-negative integer | Output token cap. |
generationConfig.candidateCount | integer | 1 | Positive integer | Number of candidates. |
generationConfig.stopSequences | array | None | ["END"] | Stop sequences. |
generationConfig.responseMimeType | string | None | application/json, text/plain | Response MIME type. |
generationConfig.responseSchema | object | None | JSON Schema | Structured output schema. |
generationConfig.presencePenalty | number | 0 | Provider-specific | Topic repetition penalty. |
generationConfig.frequencyPenalty | number | 0 | Provider-specific | Token frequency penalty. |
generationConfig.seed | integer | None | Integer | Random seed. |
generationConfig.thinkingConfig | object | Model default | {"thinkingBudget":1024} | Thinking budget or toggle. |
generationConfig.responseModalities | array | Model default | ["TEXT"], ["TEXT","IMAGE"] | Output modalities, depending on model support. |
Image generation parameters
| Parameter | Type | Default | Range / example | Meaning |
|---|
model | string | None | Image model name | Image generation model. |
prompt | string | None | "A product poster" | Image generation prompt. |
n | integer | 1 | 1 to 128 | Number of images. More images cost more. |
size | string | Model default | 1024x1024, 1024x1792 | Output size. Allowed values depend on model. |
quality | string | auto | auto, standard, hd, etc. | Image quality tier, model-dependent. |
response_format | string | Model default | url, b64_json | Image response format. |
background | string / object | Model default | transparent, opaque | Background configuration. |
output_format | string | Model default | png, jpeg, webp | Output image format. |
output_compression | integer | Model default | 0 to 100 | Compression quality, supported by some formats. |
stream | boolean | false | true / false | Stream progress or partial images where supported. |
metadata / extra fields | object | None | Provider-specific | Vendor-specific passthrough parameters. |
Embedding and rerank parameters
| Parameter | Type | Default | Range / example | Meaning |
|---|
input | string / array | None | Text or text array | Text to embed. |
model | string | None | Embedding model name | Embedding model. |
encoding_format | string | float | float, base64 | Vector encoding format. |
dimensions | integer | Model default | Positive integer | Output vector dimensions, if supported. |
query | string | None | Query text | Query for reranking. |
documents | array | None | Candidate documents | Documents to rerank. Field names may vary by model or provider. |
Audio parameters
| Parameter | Type | Default | Range / example | Meaning |
|---|
model | string | None | Audio model name | Speech model to call. |
input | string | None | Text to speak or process | TTS or audio task input. |
voice | string | Model default | alloy, nova, etc. | Voice name, depending on model support. |
instructions | string | None | Style or tone instructions | Speech generation guidance. |
response_format | string | Model default | mp3, wav, json, etc. | Response format. |
speed | number | 1 | Usually 0.25 to 4 | Speech speed, depending on model support. |
stream_format | string | None | sse | Streaming audio format. |
metadata | object | None | Custom business fields | Business tracing or vendor extensions. |
| Parameter | Type | Default | Platform behavior | Meaning |
|---|
store | boolean | Upstream default | Passed through by default; can be disabled by channel settings | Controls whether upstream may store request or output. |
safety_identifier | string | None | Filtered by default unless a channel explicitly allows it | Upstream safety identifier. Use carefully because it may identify users. |
stream_options.include_obfuscation | boolean | Upstream default | Filtered by default unless explicitly allowed | Controls streaming obfuscation behavior. |
metadata | object | None | May be logged or passed through | Custom metadata. Do not put API keys, passwords, or personal sensitive data here. |
Recommendations
- For a first call, send only
model plus messages or input.
- If you are unsure, adjust only one sampling parameter first, usually
temperature.
- Set
max_tokens or max_output_tokens when cost or response length matters.
- For tool calling, start with a simple schema before adding complex fields.
- Do not put API keys, database URLs, identity numbers, or other sensitive data in
metadata or prompts.
