Sampling parameters shape the token generation process. BazaarLink passes supported parameters to the upstream provider; unsupported parameters are silently ignored.
Pure passthrough — no local defaults
Sampling values are forwarded without range validation. The model may clamp or ignore values outside its supported range.
temperature
number
Sampling values are forwarded without range validation. The model may clamp or ignore values outside its supported range.
top_p
number
Nucleus sampling probability mass. Default: 1
top_k
integer
Sampling values are forwarded without range validation. The model may clamp or ignore values outside its supported range.
frequency_penalty
number
Penalize repeated tokens. Range: [-2, 2]. Default: 0
presence_penalty
number
Penalize tokens based on presence. Range: [-2, 2]. Default: 0
repetition_penalty
number
Reduce token repetition from input. Range: (0, 2]. Default: 1
min_p
number
Minimum probability relative to the top token. Range: [0, 1]. Default: 0
top_a
number
Dynamic top-P based on highest-probability token. Range: [0, 1]. Default: 0
seed
integer
Integer seed for deterministic sampling. Not guaranteed for all models
max_tokens
integer
Maximum number of tokens to generate
n
integer
Number of completions to generate. Default: 1
logit_bias
object
Map token IDs to bias values [-100, 100] added before sampling. OpenAI-family models only — see note below.
logprobs
boolean
Return log probabilities of each output token. OpenAI-family models only — see note below.
top_logprobs
integer
Number of most-likely tokens to return per position (requires logprobs: true). Range: 0–20. OpenAI-family models only — see note below.
response_format
object
Force structured JSON output. See Structured Output section
structured_outputs
boolean
Request strict JSON-schema-conformant output on providers that support it. Passed through as-is
reasoning
object
Provider-specific reasoning/thinking configuration. Passed through as-is
reasoning_effort
string
OpenAI o-series style reasoning effort: "low", "medium", or "high". Passed through as-is
include_reasoning
boolean
true returns the thinking text in message.reasoning (delta.reasoning when streaming); false turns thinking off for the models that think by default (see the reasoning section), so no thinking is billed. Without reasoning_effort, reasoning or thinking, only those models think and return it by default; other models return it only when asked
stop
string | string[]
Stop sequences — generation halts when encountered
tools
Tool[]
List of tools (functions) the model may call
tool_choice
string | object
Controls tool use: "auto", "none", or specific tool
parallel_tool_calls
boolean
Enable parallel function calling when tools are provided. Default: true. OpenAI-family models only — see note below.
