> ## Documentation Index > Fetch the complete documentation index at: https://docs.akumi.eu/llms.txt > Use this file to discover all available pages before exploring further. # Create a chat completion > OpenAI-compatible: point an existing OpenAI client at this base URL and change nothing else. Pass a single `model`, or an ordered `models` chain to fall back automatically when a provider is unavailable. Every request is metered and audited. ## OpenAPI ````yaml /api-reference/openapi.json post /chat/completions openapi: 3.1.0 info: title: Akumi version: 0.0.1 servers: - url: https://api.akumi.eu/v1 description: Production security: - http: [] tags: - name: Audit log - name: Chat completions - name: Embeddings - name: Models - name: Recall - name: Scores paths: /chat/completions: post: tags: - Chat completions summary: Create a chat completion description: >- OpenAI-compatible: point an existing OpenAI client at this base URL and change nothing else. Pass a single `model`, or an ordered `models` chain to fall back automatically when a provider is unavailable. Every request is metered and audited. operationId: v1.chat.completions requestBody: required: true content: application/json: schema: $ref: '#/components/schemas/ChatCompletionsRequest' responses: '200': description: '' content: application/json: schema: type: object properties: id: type: string object: type: string const: chat.completion created: type: integer model: type: string choices: type: array items: type: object properties: index: type: integer message: type: object properties: role: type: string const: assistant content: type: - string - 'null' tool_calls: type: array items: type: object properties: id: type: string type: type: string const: function function: type: object properties: name: type: string arguments: type: string required: - name - arguments required: - id - type - function required: - role - content finish_reason: type: string required: - index - message - finish_reason usage: type: object properties: prompt_tokens: type: integer completion_tokens: type: integer total_tokens: type: integer credits: type: object properties: base: type: integer services: type: integer total: type: integer required: - base - services - total required: - prompt_tokens - completion_tokens - total_tokens - credits sources: type: array items: type: object additionalProperties: {} required: - id - object - created - model - choices - usage - sources text/event-stream: schema: type: string headers: Transfer-Encoding: required: true schema: type: string enum: - chunked '401': $ref: '#/components/responses/AuthenticationException' '422': $ref: '#/components/responses/ValidationException' components: schemas: ChatCompletionsRequest: type: object properties: model: type: string description: |- Exactly one of `model` / `models` is required: a single model slug, or an ordered chain of slugs/aliases to try in turn (see RouteIntent and RoutePlanner). `prohibits:models` on `model` is what turns sending both into a 422 rather than `models` silently winning. models: type: array items: type: string minItems: 1 provider: type: object description: >- The `provider` preferences object (OpenRouter-parity routing controls). Every field may only narrow what workspace policy already permits; see App\Services\Inference\Data\ProviderPreferences. properties: order: type: - array - 'null' items: type: string only: type: - array - 'null' items: type: string ignore: type: - array - 'null' items: type: string allow_fallbacks: type: - boolean - 'null' require_parameters: type: - boolean - 'null' residency: type: - string - 'null' description: |- Validated against the accepted set here, in front of App\Services\Inference\Enums\Residency::fromWireValue(), which throws on anything it does not recognise. Without this rule an unrecognised value would reach that parser from a real request and turn into an unhandled 500 instead of a 422. enum: - eu - any - null messages: type: array items: type: object properties: role: type: string enum: - system - user - assistant - tool content: type: - string - 'null' name: type: - string - 'null' tool_call_id: type: - string - 'null' tool_calls: type: - array - 'null' items: type: object properties: id: type: string function: type: object properties: name: type: string arguments: type: - string - 'null' required: - role minItems: 1 temperature: type: - number - 'null' max_tokens: type: - integer - 'null' minimum: 1 stream: type: - boolean - 'null' stream_options: type: object description: |- OpenAI's opt-in for a trailing usage-only chunk on streamed responses. Usage is always attached to the finish chunk regardless (see RunChatCompletion::streamResponse()); this only controls the additional empty-choices frame emitted after it. properties: include_usage: type: - boolean - 'null' firewall: type: - boolean - 'null' description: |- Defaults true. Set false to opt out of the firewall and accept non-EU routing; honoured only when the organization has recorded a compliance acceptance, otherwise the egress guard refuses the call. firewall_language: anyOf: - $ref: '#/components/schemas/PiiLanguage' description: |- The language the firewall analyses this request as, overriding the workspace's setting for this call only. `auto` detects per message. A caller sending several languages needs this per call rather than a workspace each; one that always sends the same language should set it on the workspace and omit it here. - type: 'null' collection: type: - array - 'null' description: |- Narrows retrieval to named knowledge collections. Omit it (the default) to blend the whole workspace: personal memory, the workspace's shared knowledge, and every collection in it. Pass a collection slug, or a list of slugs, to read only those plus the caller's own personal memory. A slug the workspace does not have is rejected, never ignored. items: type: string maxLength: 255 user: type: - string - 'null' description: >- Optional end-user id (OpenAI's `user`). Scopes long-term memory facts. maxLength: 255 thread: type: - string - 'null' description: >- Optional conversation thread id. Present = opt into server-side thread replay/append (Memory service); a client-chosen, organization-scoped string. maxLength: 255 tags: type: - array - 'null' description: |- Free-form labels recorded on the trace, at most ten of at most 64 characters each. They do not change how the request is served: they make the trace explorer able to answer questions about it afterwards. items: type: string maxLength: 64 maxItems: 10 name: type: - string - 'null' description: A human label for this request, shown on its trace. maxLength: 120 top_p: type: - number - 'null' minimum: 0 maximum: 1 stop: type: - array - 'null' description: >- stop accepts a bare string or an array of strings; `stop.*` validates array elements when an array is supplied but does not apply to a scalar. items: type: string presence_penalty: type: - number - 'null' minimum: -2 maximum: 2 frequency_penalty: type: - number - 'null' minimum: -2 maximum: 2 'n': type: - integer - 'null' minimum: 1 seed: type: - integer - 'null' logit_bias: type: - array - 'null' items: type: number minimum: -100 maximum: 100 logprobs: type: - boolean - 'null' top_logprobs: type: - integer - 'null' minimum: 0 maximum: 20 max_completion_tokens: type: - integer - 'null' minimum: 1 tools: type: - array - 'null' items: type: object properties: type: type: - string - 'null' enum: - function - null function: type: object properties: name: type: string description: type: - string - 'null' parameters: type: - array - 'null' items: type: string tool_choice: type: - string - 'null' description: >- tool_choice may be a string ('none'|'auto'|'required') OR an object ({type:'function',function:{name}}); the controller normalises it downstream. parallel_tool_calls: type: - boolean - 'null' cache: type: - boolean - 'null' description: |- Set false to bypass the response cache for this request even when the Cache service is enabled and the request would otherwise be cacheable. Defaults true (cache enabled when service is on). response_format: type: object properties: type: type: string enum: - text - json_object - json_schema json_schema: type: - object - 'null' properties: name: type: string schema: type: - array - 'null' items: type: string strict: type: - boolean - 'null' required: - messages title: ChatCompletionsRequest PiiLanguage: type: string description: > The language a workspace's text is analysed as. WHY THIS EXISTS AT ALL. The self-hosted GLiNER detector is one multilingual model that reads Dutch, German and French without being told which, so the platform never had to know. Azure AI Language does have to be told: it rejects `auto` outright, and analysing Dutch as English is not a no-op. A Dutch citizen number comes back as `EUNationalIdentificationNumber` instead of `NLCitizensServiceNumber`, which the mapping does not know, so the BSN goes out in plaintext. Getting this wrong leaks. ONLY THE FOUR THE CORPORA COVER. Azure supports around eighty languages, and offering them here would claim a detection quality nobody has measured. These four are the ones `pii-corpus-gretel.json` and `pii-eval-cases.json` score, so these four are the ones offered. Adding a fifth means adding cases for it first. enum: - auto - nl - en - fr - de title: PiiLanguage responses: AuthenticationException: description: Unauthenticated content: application/json: schema: type: object properties: message: type: string description: Error overview. required: - message ValidationException: description: Validation error content: application/json: schema: type: object properties: message: type: string description: Errors overview. errors: type: object description: A detailed description of each field that failed validation. additionalProperties: type: array items: type: string required: - message - errors securitySchemes: http: type: http scheme: bearer ````