Créer une réponse de modèle
Utilisez CometAPI POST /v1/responses pour créer des réponses de modèle multimodales et un historique de conversation avec des outils intégrés et function calling.
gpt-6-astra. Les exemples montrent également la recherche web, les appels de fonction personnalisés, les sorties structurées, le raisonnement et le Streaming.
Poursuivre une conversation
Pour une conversation uniquement textuelle, conservez les messages utilisateur et les réponses textuelles de l’assistant, puis ajoutez le message utilisateur suivant :response.output_text est une propriété pratique du SDK qui combine le texte de l’assistant provenant de la réponse. Les conversations utilisant function calling nécessitent également leurs éléments d’appel d’outil et de résultat, comme décrit ci-dessous.
Utiliser la recherche web
Ajoutez un outilweb_search pour rechercher des informations et inclure des citations de sources :
web_search_call renvoyés et les annotations d’URL du message lorsque votre application a besoin de résultats de recherche ou de citations. Les autres outils intégrés nécessitent leur propre configuration ; consultez le guide des outils OpenAI.
Appeler des fonctions personnalisées
Définissez les arguments de la fonction et laissez le modèle demander l’appel de la fonction :function_call. Son champ arguments est une chaîne encodée en JSON. Exécutez votre fonction, puis construisez la requête suivante à partir de l’entrée d’origine et des éléments de sortie renvoyés, sans leurs champs id d’élément de réponse. Conservez les name, arguments et call_id de la fonction, puis ajoutez un élément function_call_output avec le call_id correspondant et votre résultat dans output.
Demander une sortie structurée
Utiliseztext.format pour demander du JSON conforme à un schéma :
Configurer le raisonnement
Définissezreasoning.effort pour sélectionner l’effort de raisonnement du modèle :
low pour les tâches générales et high pour l’exemple de raisonnement. max_output_tokens inclut à la fois les tokens de raisonnement et les tokens de sortie visibles.
Diffuser les réponses en continu
Définissezstream sur true pour recevoir des événements envoyés par le serveur :
response.created, response.in_progress et response.completed. Le texte arrive dans des événements response.output_text.delta. Les événements d’outil et de raisonnement dépendent de la requête ; traitez donc les événements selon leur type plutôt que de supposer que chaque réponse suit la même séquence.
Autorisations
Bearer token authentication. Use your CometAPI key.
Corps
Model ID to use for this request. See the Models page for current options.
"gpt-6-astra"
Text, image, or file inputs to the model, used to generate a response. Can be a simple string for text-only input, or an array of input items for multimodal content (images, files) and multi-turn conversations.
A system (or developer) message inserted into the model's context. When used with previous_response_id, instructions from the previous response are not carried over — this makes it easy to swap system messages between turns.
Whether to run the model response in the background. Background responses do not return output directly — you retrieve the result later via the response ID.
Context management configuration for this request. Controls how the model manages context when the conversation exceeds the context window.
The conversation this response belongs to. Items from the conversation are prepended to input for context. Input and output items are automatically added to the conversation after the response completes. Cannot be used with previous_response_id.
Additional output data to include in the response. Use this to request extra information that is not included by default.
web_search_call.action.sources, code_interpreter_call.outputs, computer_call_output.output.image_url, file_search_call.results, message.input_image.image_url, message.output_text.logprobs, reasoning.encrypted_content An upper bound for the number of tokens that can be generated for a response, including visible output tokens and reasoning tokens.
The maximum number of total calls to built-in tools that can be processed in a response. This limit applies across all built-in tool calls, not per individual tool. Any further tool call attempts by the model will be ignored.
Set of up to 16 key-value pairs that can be attached to the response. Useful for storing additional information in a structured format. Keys have a maximum length of 64 characters; values have a maximum length of 512 characters.
Whether to allow the model to run tool calls in parallel.
An ID for a stored response that is accessible to the selected model and request context. Cannot be combined with conversation. The conversation-history example sends prior input and output items explicitly.
Reference to a prompt template and its variables.
A key used to cache responses for similar requests, helping optimize cache hit rates. Replaces the deprecated user field for caching purposes.
Legacy prompt-cache retention setting. Refer to the selected model documentation for prompt_cache_options and supported retention settings.
in_memory, 24h Reasoning configuration for the selected model.
A stable identifier for your end-users, used to help detect policy violations. Should be a hashed username or email — do not send identifying information directly.
64Specifies the processing tier for the request. When set, the response will include the actual service_tier used.
auto: Uses the tier configured in project settings (default behavior).default: Standard pricing and performance.flex: Flexible processing with potential cost savings.priority: Priority processing with faster response times.
auto, default, flex, priority Request response storage for later retrieval, when supported by the selected model and request context.
If set to true, the response data will be streamed to the client as it is generated using server-sent events (SSE). Events include response.created, response.output_text.delta, response.completed, and more.
Options for streaming responses. Only set this when stream is true.
Sampling temperature. Omit this field for GPT-6 Astra.
0 <= x <= 2Configuration for text output. Use this to request structured JSON output via JSON mode or JSON Schema.
Controls how the model selects which tool(s) to call.
auto(default): The model decides whether and which tools to call.none: The model will not call any tools.required: The model must call at least one tool.- An object specifying a particular tool to use.
Tools available to the model. This page demonstrates web_search and custom function tools. Other tool types require their own configuration; see the official tool reference.
Number of token alternatives to include for models supporting log probabilities. Requires message.output_text.logprobs in include. Omit for GPT-6 Astra.
0 <= x <= 20Nucleus sampling threshold. Omit this field for GPT-6 Astra. For sampling overrides, adjust either top_p or temperature.
0 <= x <= 1The truncation strategy for handling inputs that exceed the model's context window.
auto: The model truncates the input by dropping items from the beginning of the conversation to fit.disabled(default): The request fails with a 400 error if the input exceeds the context window.
auto, disabled Deprecated. Use safety_identifier and prompt_cache_key instead. A stable identifier for your end-user.
Réponse
The generated Response object.
Unique identifier for the response.
"resp_example"
The object type, always response.
response "response"
Unix timestamp (in seconds) of when the response was created.
1788763707
The status of the response.
completed, in_progress, failed, cancelled, queued, incomplete "completed"
Whether the response was run in the background.
false
Unix timestamp of when the response was completed, or null if still in progress.
1788763711
Error information if the response failed, or null on success.
Details about why the response is incomplete, if applicable.
The system instructions used for this response.
The maximum output token limit that was applied.
The model used for the response.
"gpt-6-astra"
An array of output items generated by the model. Each item can be a message, function call, or other output type.
Whether parallel tool calls were enabled.
The ID of the previous response, if this is a multi-turn conversation.
The reasoning configuration that was used.
The service tier actually used to process the request.
Whether the response was stored.
The temperature value used.
The text configuration used.
The tool choice setting used.
The tools that were available for this response.
The top_p value used.
The truncation strategy used.
Token usage statistics for this response.
The user identifier, if provided.
The metadata attached to this response.
Content filter results applied to the response, if any.
The frequency penalty applied to the request.
Maximum number of tool calls allowed, if set.
The presence penalty applied to the request.
Cache key for prompt caching, if applicable.
Prompt cache retention policy, if applicable.
Safety system identifier for the response, if applicable.
Number of top log probabilities returned per token position.