asyncapi: '2.6.0'
info:
  title: DeepSeek Streaming API (HTTP + SSE)
  version: '1.0.0'
  description: |
    AsyncAPI definition for the streaming surface of the DeepSeek API.

    DeepSeek exposes an OpenAI-compatible HTTP API. When the `stream` request
    parameter is set to `true`, the server upgrades the response to a
    `text/event-stream` (Server-Sent Events) channel and emits a sequence of
    data-only events. Each event contains a JSON payload representing a
    streaming chunk, and the stream is terminated by a sentinel `data: [DONE]`
    event.

    Two streaming surfaces are described:

      * `/chat/completions` (production) - chat completions for `deepseek-chat`
        and `deepseek-reasoner`. The `deepseek-reasoner` model additionally
        emits a `reasoning_content` field in `delta` chunks during the
        chain-of-thought phase, followed by `content` deltas for the final
        answer.
      * `/beta/completions` (beta) - Fill-In-the-Middle (FIM) text completions
        that emit `text_completion.chunk` events using the OpenAI legacy
        completions stream shape.

    Only events documented at https://api-docs.deepseek.com are represented;
    no fields have been inferred or fabricated beyond the official docs.
  contact:
    name: DeepSeek API Docs
    url: https://api-docs.deepseek.com
  license:
    name: DeepSeek Terms of Use
    url: https://chat.deepseek.com/downloads/DeepSeek%20Terms%20of%20Use.html

defaultContentType: text/event-stream

servers:
  production:
    url: https://api.deepseek.com
    protocol: https
    description: |
      DeepSeek OpenAI-compatible HTTPS endpoint. Streaming responses are
      delivered as `text/event-stream` (Server-Sent Events) when the request
      body sets `"stream": true`.
    security:
      - bearerAuth: []
    bindings:
      http:
        bindingVersion: '0.3.0'
  beta:
    url: https://api.deepseek.com/beta
    protocol: https
    description: |
      DeepSeek beta HTTPS endpoint. Required base URL for the FIM
      (Fill-In-the-Middle) completions API. Streaming responses are delivered
      as `text/event-stream` (Server-Sent Events) when the request body sets
      `"stream": true`.
    security:
      - bearerAuth: []
    bindings:
      http:
        bindingVersion: '0.3.0'

channels:
  /chat/completions:
    description: |
      Server-Sent Events stream of `chat.completion.chunk` objects produced
      by `POST /chat/completions` when `stream` is `true`. The first chunk
      typically carries `delta.role = "assistant"`. Subsequent chunks carry
      incremental `delta.content` tokens, and - for the `deepseek-reasoner`
      model - `delta.reasoning_content` tokens during the chain-of-thought
      phase. The final data chunk before the terminator may include a
      populated `usage` object when `stream_options.include_usage` is set to
      `true`. The stream is closed by a literal `data: [DONE]` event.
    servers:
      - production
    bindings:
      http:
        bindingVersion: '0.3.0'
        type: response
        method: POST
    subscribe:
      operationId: subscribeChatCompletionStream
      summary: Receive streaming chat completion chunks.
      description: |
        Each Server-Sent Event has the form `data: <json>` where `<json>` is
        either a `chat.completion.chunk` object or the literal string `[DONE]`
        used as the stream terminator.
      bindings:
        http:
          bindingVersion: '0.3.0'
      message:
        oneOf:
          - $ref: '#/components/messages/ChatCompletionChunk'
          - $ref: '#/components/messages/StreamDone'

  /beta/completions:
    description: |
      Server-Sent Events stream of `text_completion` chunks produced by
      `POST /beta/completions` (the Fill-In-the-Middle completions endpoint)
      when `stream` is `true`. Each event is a partial completion chunk
      following the OpenAI legacy completions streaming shape. The stream is
      closed by a literal `data: [DONE]` event.
    servers:
      - beta
    bindings:
      http:
        bindingVersion: '0.3.0'
        type: response
        method: POST
    subscribe:
      operationId: subscribeFimCompletionStream
      summary: Receive streaming FIM (Fill-In-the-Middle) completion chunks.
      description: |
        Each Server-Sent Event has the form `data: <json>` where `<json>` is
        either a `text_completion` streaming chunk or the literal string
        `[DONE]` used as the stream terminator.
      bindings:
        http:
          bindingVersion: '0.3.0'
      message:
        oneOf:
          - $ref: '#/components/messages/FimCompletionChunk'
          - $ref: '#/components/messages/StreamDone'

components:
  securitySchemes:
    bearerAuth:
      type: http
      scheme: bearer
      bearerFormat: API Key
      description: |
        DeepSeek API key passed as `Authorization: Bearer <DEEPSEEK_API_KEY>`
        on the originating HTTP request that opens the SSE stream.

  messages:
    ChatCompletionChunk:
      name: ChatCompletionChunk
      title: Chat Completion Streaming Chunk
      summary: |
        A single `chat.completion.chunk` event emitted on the SSE stream while
        a `/chat/completions` request with `stream=true` is in progress.
      contentType: application/json
      bindings:
        http:
          bindingVersion: '0.3.0'
      payload:
        $ref: '#/components/schemas/ChatCompletionChunk'
      examples:
        - name: ChatRoleChunk
          summary: First delta in a stream, announcing the assistant role.
          payload:
            id: chatcmpl-abc123
            object: chat.completion.chunk
            created: 1748563200
            model: deepseek-chat
            system_fingerprint: fp_44709d6fcb
            choices:
              - index: 0
                delta:
                  role: assistant
                  content: ''
                finish_reason: null
        - name: ChatContentDelta
          summary: Incremental content token from `deepseek-chat`.
          payload:
            id: chatcmpl-abc123
            object: chat.completion.chunk
            created: 1748563200
            model: deepseek-chat
            system_fingerprint: fp_44709d6fcb
            choices:
              - index: 0
                delta:
                  content: 'Hello'
                finish_reason: null
        - name: ReasonerReasoningDelta
          summary: Chain-of-thought delta from `deepseek-reasoner`.
          payload:
            id: chatcmpl-xyz789
            object: chat.completion.chunk
            created: 1748563205
            model: deepseek-reasoner
            system_fingerprint: fp_44709d6fcb
            choices:
              - index: 0
                delta:
                  reasoning_content: 'Let me think step by step.'
                finish_reason: null
        - name: ReasonerFinalContent
          summary: Final answer delta after reasoning has completed.
          payload:
            id: chatcmpl-xyz789
            object: chat.completion.chunk
            created: 1748563205
            model: deepseek-reasoner
            system_fingerprint: fp_44709d6fcb
            choices:
              - index: 0
                delta:
                  content: 'The answer is 42.'
                finish_reason: null
        - name: ChatFinalChunkWithUsage
          summary: Terminating chunk with `finish_reason` and usage details.
          payload:
            id: chatcmpl-abc123
            object: chat.completion.chunk
            created: 1748563200
            model: deepseek-chat
            system_fingerprint: fp_44709d6fcb
            choices:
              - index: 0
                delta: {}
                finish_reason: stop
            usage:
              prompt_tokens: 12
              completion_tokens: 24
              total_tokens: 36
              prompt_cache_hit_tokens: 0
              prompt_cache_miss_tokens: 12

    FimCompletionChunk:
      name: FimCompletionChunk
      title: FIM Completion Streaming Chunk
      summary: |
        A single streaming chunk emitted on the SSE stream while a
        `/beta/completions` (Fill-In-the-Middle) request with `stream=true`
        is in progress.
      contentType: application/json
      bindings:
        http:
          bindingVersion: '0.3.0'
      payload:
        $ref: '#/components/schemas/FimCompletionChunk'
      examples:
        - name: FimTextDelta
          summary: Incremental text fragment from a FIM completion.
          payload:
            id: cmpl-fim-abc
            object: text_completion
            created: 1748563300
            model: deepseek-chat
            choices:
              - index: 0
                text: '    return a + b'
                finish_reason: null
                logprobs: null
        - name: FimFinalChunkWithUsage
          summary: Terminating chunk with `finish_reason` and usage details.
          payload:
            id: cmpl-fim-abc
            object: text_completion
            created: 1748563300
            model: deepseek-chat
            choices:
              - index: 0
                text: ''
                finish_reason: stop
                logprobs: null
            usage:
              prompt_tokens: 8
              completion_tokens: 16
              total_tokens: 24
              prompt_cache_hit_tokens: 0
              prompt_cache_miss_tokens: 8

    StreamDone:
      name: StreamDone
      title: SSE Stream Terminator
      summary: |
        Sentinel event marking the end of the SSE stream. The raw SSE line
        is `data: [DONE]`. After this event the server closes the response
        body.
      contentType: text/plain
      bindings:
        http:
          bindingVersion: '0.3.0'
      payload:
        $ref: '#/components/schemas/StreamDone'
      examples:
        - name: Done
          summary: Stream terminator emitted after all data chunks.
          payload: '[DONE]'

  schemas:
    ChatCompletionChunk:
      type: object
      description: |
        A streaming chunk of a chat completion response. Returned by
        `POST /chat/completions` when `stream` is set to `true`. The shape
        mirrors the OpenAI Chat Completions streaming chunk and adds DeepSeek
        specific fields (`reasoning_content` for `deepseek-reasoner`, and
        prompt cache hit / miss counters on the terminal `usage` object).
      required:
        - id
        - object
        - created
        - model
        - choices
      properties:
        id:
          type: string
          description: |
            Unique identifier for the chat completion. The same `id` is shared
            across every chunk of a single streamed response.
        object:
          type: string
          enum:
            - chat.completion.chunk
          description: Object type. Always `chat.completion.chunk` for streamed events.
        created:
          type: integer
          format: int64
          description: |
            Unix timestamp (seconds) when the completion was created. Identical
            across every chunk of a single streamed response.
        model:
          type: string
          description: |
            Identifier of the model that produced the chunk
            (for example `deepseek-chat` or `deepseek-reasoner`).
        system_fingerprint:
          type: string
          description: Backend configuration fingerprint.
        choices:
          type: array
          description: Array of streamed choice deltas.
          items:
            $ref: '#/components/schemas/ChatCompletionChunkChoice'
        usage:
          description: |
            Token usage details. Only populated on the terminal data chunk
            (immediately before `[DONE]`) when the request was sent with
            `stream_options.include_usage = true`. Null on intermediate chunks.
          oneOf:
            - type: 'null'
            - $ref: '#/components/schemas/ChatCompletionUsage'

    ChatCompletionChunkChoice:
      type: object
      required:
        - index
        - delta
      properties:
        index:
          type: integer
          description: Index of this choice in the `choices` array.
        delta:
          $ref: '#/components/schemas/ChatCompletionDelta'
        finish_reason:
          description: |
            Reason the model stopped generating tokens for this choice. Null
            on all chunks except the final delta for the choice.
          oneOf:
            - type: 'null'
            - type: string
              enum:
                - stop
                - length
                - content_filter
                - tool_calls
                - insufficient_system_resource
        logprobs:
          description: |
            Log-probability information for the streamed tokens. Present only
            when the originating request specified `logprobs: true`.
          oneOf:
            - type: 'null'
            - type: object

    ChatCompletionDelta:
      type: object
      description: |
        Incremental update applied to the assistant message for this choice.
        Fields are only present on chunks that contribute new information.
      properties:
        role:
          type: string
          enum:
            - assistant
          description: |
            Role of the streamed message. Emitted on the first delta of a
            streamed response.
        content:
          type: string
          nullable: true
          description: |
            Incremental fragment of the final assistant content. For
            `deepseek-reasoner` this is emitted only after the
            `reasoning_content` phase completes.
        reasoning_content:
          type: string
          nullable: true
          description: |
            Incremental fragment of the assistant's Chain of Thought. Emitted
            only by the `deepseek-reasoner` model during its reasoning phase.
            This field is output-only - including `reasoning_content` in a
            subsequent request's input messages returns a 400 error.
        tool_calls:
          type: array
          description: |
            Incremental tool call fragments. Each entry mirrors the
            non-streamed `tool_calls` shape (`id`, `type: "function"`,
            `function: { name, arguments }`) with `arguments` streamed as a
            growing JSON string.
          items:
            type: object

    ChatCompletionUsage:
      type: object
      description: Token usage details for the completed streamed response.
      required:
        - prompt_tokens
        - completion_tokens
        - total_tokens
      properties:
        prompt_tokens:
          type: integer
          description: Number of tokens in the prompt.
        completion_tokens:
          type: integer
          description: Number of tokens in the generated completion.
        total_tokens:
          type: integer
          description: Sum of `prompt_tokens` and `completion_tokens`.
        prompt_cache_hit_tokens:
          type: integer
          description: |
            Number of prompt tokens served from DeepSeek's context cache.
        prompt_cache_miss_tokens:
          type: integer
          description: |
            Number of prompt tokens that missed the context cache and were
            processed fresh.
        completion_tokens_details:
          type: object
          description: Detailed breakdown of completion tokens.
          properties:
            reasoning_tokens:
              type: integer
              description: |
                Number of tokens consumed by the model's reasoning
                (Chain of Thought) phase. Populated for `deepseek-reasoner`.

    FimCompletionChunk:
      type: object
      description: |
        Streaming chunk for the Fill-In-the-Middle (FIM) completions endpoint
        at `POST /beta/completions`. Follows the OpenAI legacy completions
        stream shape.
      required:
        - id
        - object
        - created
        - model
        - choices
      properties:
        id:
          type: string
          description: Unique identifier for the FIM completion.
        object:
          type: string
          enum:
            - text_completion
          description: Object type. Always `text_completion` for FIM stream chunks.
        created:
          type: integer
          format: int64
          description: Unix timestamp (seconds) when the completion was created.
        model:
          type: string
          description: Identifier of the model that produced the chunk.
        choices:
          type: array
          description: Array of streamed FIM completion choices.
          items:
            $ref: '#/components/schemas/FimCompletionChunkChoice'
        usage:
          description: |
            Token usage details. Only populated on the terminal data chunk
            (immediately before `[DONE]`) when the request was sent with
            `stream_options.include_usage = true`. Null on intermediate chunks.
          oneOf:
            - type: 'null'
            - $ref: '#/components/schemas/ChatCompletionUsage'

    FimCompletionChunkChoice:
      type: object
      required:
        - index
        - text
      properties:
        index:
          type: integer
          description: Index of this choice in the `choices` array.
        text:
          type: string
          description: Incremental text fragment generated by the model.
        finish_reason:
          description: |
            Reason the model stopped generating tokens for this choice. Null
            on all chunks except the final delta for the choice.
          oneOf:
            - type: 'null'
            - type: string
              enum:
                - stop
                - length
                - content_filter
                - insufficient_system_resource
        logprobs:
          description: |
            Token log-probability information. Present only when the
            originating request specified `logprobs`.
          oneOf:
            - type: 'null'
            - type: object

    StreamDone:
      type: string
      description: |
        Literal string `[DONE]` emitted as the final SSE data line. Indicates
        the stream is closed and no further chunks will follow.
      enum:
        - '[DONE]'