Create chat completion

POST

/chat/completions

curl https://api.waibee.com/v1/chat/completions \
  -H "Authorization: Bearer $WAIBEE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4.6",
    "messages": [{ "role": "user", "content": "Hello!" }]
  }'

OpenAI-compatible chat completion. Send messages, receive a model response. Set stream: true for a Server-Sent Events stream terminated by data: [DONE].

Authorizations

• bearerAuth

Request Body^required

Media typeapplication/json

object

model

required

<provider>/<model> from GET /v1/models. waibee/auto lets the router pick.

string

messages

required

Array<object>

>= 1 items

object

role

required

string

Allowed values: system user assistant

content

required

string

stream

Stream the response as Server-Sent Events.

boolean

max_tokens

Maximum tokens to generate. max_completion_tokens is also accepted.

integer

temperature

number

<= 2

top_p

number

tools

Function/tool definitions (max 256).

Array<object>

Function/tool definition for function calling.

object

type

string

Allowed value: function

function

object

name

string

description

string

parameters

JSON Schema of the tool arguments.

object

key

additional properties

any

tool_choice

Tool selection: auto | none | required, or an object {type: function, function: {name}}.

string

web_search_options

Ask the router to run a web search for this request. An empty object is enough. Each executed search is billed on top of tokens, and the sources come back as message.annotations[].url_citation.

object

key

additional properties

any

response_format

Structured output format.

object

type

string

Allowed values: text json_object json_schema

json_schema

Schema used when type is json_schema.

object

key

additional properties

any

reasoning_effort

Reasoning depth. max is NOT supported here, use /v1/messages with output_config.effort.

string

Allowed values: minimal low medium high xhigh

plugins

Per-request plugins.

object

dlp

Data-loss-prevention scan, mask or block. Optional: a DLP policy set for your API key applies as a floor of strictness even when this object is absent, and this object can only make the check stricter. entity_types is unioned with the policy’s categories, so the scan may be wider than the list you send.

object

mode

Applies to the rules you declare yourself. Where the key’s policy declares the same rule, the stricter of the two modes wins; the policy’s own categories keep the mode the administrator set.

string

Allowed values: alert mask block

entity_types

Unioned with the categories of the key’s policy, never narrowing them.

Array<string>

report_scope

Which messages’ findings come back to you. Narrowing it does not narrow the scan: every message is still scanned, and the block decision counts every finding.

string

default: all

Allowed values: all last_message latest_user_message

stop_words

Values that are a finding on their own, whatever the categories say. Unioned with the key’s policy.

Array<object>

object

value

required

string

match

string

default: exact

Allowed values: exact fuzzy

allowlist

Exceptions, matched on the whole value only. They reach your own rules only: an exception never lifts a finding raised by a rule of the key’s policy. Within your own config a stop word beats an exception on the same value.

Array<object>

object

value

required

string

match

Exact only. A substring exception would exempt every lookalike containing it.

string

default: exact

Allowed values: exact

regexes

Named patterns, run by RE2. Unioned with the key’s policy.

Array<object>

object

name

required

string

pattern

required

string

injected_context_names

Marks a message as inserted by the platform rather than typed by a person, by the value of its name field, and takes it out of the scan. Ignored while a policy applies to your key: on this path the same caller supplies both the marker and the content, so it would be a way to opt out of the check.

Array<string>

acknowledged

Reserved, and currently changes nothing: alert reports the findings and forwards the request, so there is no hold to confirm.

boolean

Responses

200

Model response.

Media typeapplication/json

object

id

string

object

string

Allowed value: chat.completion

created

integer

model

string

choices

Array<object>

object

index

integer

finish_reason

string

Allowed values: stop length tool_calls content_filter

message

object

role

required

string

Allowed values: system user assistant

content

required

string

usage

object

prompt_tokens

integer

completion_tokens

integer

total_tokens

integer

prompt_tokens_details

object

cached_tokens

integer

cache_write_tokens

integer

Example

{
  "id": "gen-...",
  "object": "chat.completion",
  "created": 1784025392,
  "model": "anthropic/claude-sonnet-4.6",
  "choices": [
    {
      "index": 0,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "Paris"
      }
    }
  ],
  "usage": {
    "prompt_tokens": 22,
    "completion_tokens": 6,
    "total_tokens": 28
  }
}

400

Invalid request. Codes: invalid_model, validation_error, feature_not_supported.

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "invalid_model",
    "message": "Model 'foo' is not available",
    "extra": {}
  }
}

401

Missing or invalid API key.

Media typeapplication/json

object

error

string

Example

{
  "error": "unauthorized"
}

402

Not enough credits for the request (insufficient_funds).

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "insufficient_funds",
    "message": "Insufficient credits",
    "extra": {}
  }
}

403

Request blocked by an access policy (policy_violation) or by DLP (dlp_violation). A DLP block is about the content of the request, not about the key’s permissions: branch on error.code, not on the status alone.

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Examples

Model not allowed for this group

{
  "error": {
    "code": "policy_violation",
    "message": "Request blocked by policy",
    "extra": {}
  }
}

Sensitive data detected in block mode

{
  "error": {
    "code": "dlp_violation",
    "message": "Request blocked: sensitive data detected",
    "extra": {
      "dlp": [
        {
          "category": "pii",
          "entity_type": "EMAIL",
          "message_index": 0,
          "location": "messages[0].content",
          "start": 12,
          "end": 28,
          "confidence": 0.98
        }
      ],
      "findings_count": 1,
      "block_scope": "in_report_scope"
    }
  }
}

404

Model not in the catalog (model_not_found).

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "model_not_found",
    "message": "Model 'nonexistent/model' is not available. Check the model name and try again.",
    "extra": {}
  }
}

408

Upstream took too long, including a stream that went silent (provider_timeout).

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "provider_timeout",
    "message": "Provider timed out",
    "extra": {}
  }
}

413

Request body, a file or the aggregate text exceeds the limit: payload_too_large, file_too_large, text_content_too_large.

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "payload_too_large",
    "message": "Request body too large",
    "extra": {}
  }
}

422

Request failed validation (validation_error): model not in <provider>/<model> form, or a missing field.

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "validation_error",
    "message": "Invalid request",
    "extra": {
      "errors": [
        {
          "type": "value_error",
          "loc": [
            "body",
            "model"
          ],
          "msg": "model must be in format <provider>/<model>"
        }
      ]
    }
  }
}

429

Rate limit exceeded (provider_rate_limited).

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "provider_rate_limited",
    "message": "Rate limit exceeded",
    "extra": {}
  }
}

502

Upstream model unavailable (provider_unavailable).

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Example

{
  "error": {
    "code": "provider_unavailable",
    "message": "Provider unavailable",
    "extra": {}
  }
}

503

Temporarily unavailable. service_unavailable and circuit_breaker_open mean load or overload protection; dlp_policy_unavailable means the DLP policy for the key could not be resolved, and the request was refused rather than run unscanned. service_unavailable and dlp_policy_unavailable carry Retry-After; circuit_breaker_open does not, because the router does not know when the backend it fenced off will be back. dlp_unavailable also lands here and carries no header either: it says the router is not enforcing DLP, so a retry returns the same answer.

Media typeapplication/json

OpenAI-format error envelope.

object

error

required

object

code

string

message

string

extra

Additional error data; fields depend on the error code.

object

key

additional properties

any

Examples

Router under load

{
  "error": {
    "code": "service_unavailable",
    "message": "Service temporarily unavailable due to high load",
    "extra": {}
  }
}

DLP policy could not be resolved (retryable)

{
  "error": {
    "code": "dlp_policy_unavailable",
    "message": "DLP policy is temporarily unavailable; retry the request",
    "extra": {}
  }
}

plugins.dlp sent where DLP is not enforced

{
  "error": {
    "code": "dlp_unavailable",
    "message": "DLP plugin config was provided, but router DLP enforcement is disabled",
    "extra": {}
  }
}

Headers

Retry-After

integer

Seconds to wait before retrying.