Parameter compatibility
Providers disagree about which request parameters they accept. Prism adjusts
parameters so a request reaches the model you chose, without changing the
model, and reports every adjustment in the X-Prism-Adjusted-Params response
header.
Built-in adjustments
Section titled “Built-in adjustments”On Chat Completions:
max_tokensis renamed tomax_completion_tokensfor the OpenAI o1, o3, o4 and gpt-5 families.- A non-default
temperatureortop_pis dropped for reasoning models that accept only1(not for gpt-5-chat variants). The value1is kept. - For automatic routing only,
reasoning_effort(Responses:reasoning) is removed when the chosen model is a non-reasoning gpt-4o, gpt-4.1, gpt-4 or gpt-3.5 model.
These adjustments also apply to requests that name a concrete model.
Learned adjustments
Section titled “Learned adjustments”When a provider answers 400 or 422 and names a parameter of the request
that it does not support, Prism retries the same provider and model with the
parameter changed:
- set to the value the provider suggests, if it suggests one;
- renamed to the field the provider names, if it names one;
- otherwise dropped.
At most four such retries happen per attempt. A repaired rejection does not count as a failure of the model. For streaming chat the retry happens before any response bytes are sent; Prism never repairs a stream in the middle.
Prism remembers the fix for that provider, model and task type and applies it
to later matching requests before sending them. A rule expires seven days
after it was learned, so the original parameter is tried again; rules survive
a restart. They are stored in <storage.dir>/learned-param-rules.json, at most
16 per provider, model and task type.
Learned rules never change model, messages, input, instructions,
prompt, tools, functions, tool_choice, function_call,
response_format, text, stream, n, modalities, audio,
previous_response_id or conversation. Other wire parameters, including
reasoning_effort, can change: naming a concrete model does not guarantee the
request body reaches the provider byte for byte.
List and remove learned rules with the list_learned_rules and
forget_learned_rules actions of the local control API,
or in the web console.
Chat requests served through the Responses API
Section titled “Chat requests served through the Responses API”Some models answer only on the Responses API. When a provider’s Chat endpoint
rejects a request and points to /v1/responses, or the catalog marks a model
as Responses-only on a provider that offers that API, Prism sends the chat
request to /v1/responses and translates the answer back to Chat Completions,
streamed or not. You keep calling /v1/chat/completions.
X-Prism-Adjusted-Paramscontainsendpoint=set:responses, andX-Prism-Upstream-Endpoint: /v1/responsesis set.- Chat-only fields are dropped and listed:
stop,seed,presence_penalty,frequency_penalty,logit_bias,logprobs,top_logprobs,prediction,web_search_options. - Requests this path cannot carry stay on Chat:
ngreater than 1, audio or other non-text output, non-function tools, unsupported message roles, non-text system or developer content. For a Responses-only model such a request gets a400that names/v1/responses, and automatic routing does not pick that model for it. - If the provider’s
/v1/responsesanswers404,405or501, Prism goes back to Chat for that model for seven days.
Reading the header
Section titled “Reading the header”X-Prism-Adjusted-Params: temperature=dropped;max_tokens=renamed:max_completion_tokensEntries are separated by ;: param=dropped, param=renamed:<field>,
param=set:<value>. Browser JavaScript on another origin cannot read this
header (it is not in Access-Control-Expose-Headers); server-side clients and
same-origin pages can.