chatTurnsFromMessages function

List<LlamaChatTurn> chatTurnsFromMessages(Iterable<ChatMessage> messages)

Builds the engine-neutral structured view of messages passed to LlamaSession.generate alongside the rendered prompt for image turns.

Content parts keep the author's ordering (text and media interleaved as written). Media is extracted with the same rule the ChatFormat templates use — DataContent with non-null bytes — typed by kind so the message-level multimodal path can label each part (image vs audio). Other content kinds (tool calls, tool results) have no part representation and are dropped; their text, if any, is what reaches the model through the rendered prompt.

Implementation

List<LlamaChatTurn> chatTurnsFromMessages(Iterable<ChatMessage> messages) =>
    <LlamaChatTurn>[
      for (final message in messages)
        LlamaChatTurn(
          role: message.role == ChatRole.system
              ? LlamaChatRole.system
              : message.role == ChatRole.assistant
              ? LlamaChatRole.assistant
              : message.role == ChatRole.tool
              ? LlamaChatRole.tool
              : LlamaChatRole.user,
          parts: <LlamaContentPart>[
            for (final content in message.contents)
              if (content is TextContent)
                LlamaTextPart(content.text)
              else if (content is DataContent && content.data != null)
                if (content.hasTopLevelMediaType('image'))
                  LlamaImagePart(content.data!, mimeType: content.mediaType)
                else if (content.hasTopLevelMediaType('audio'))
                  LlamaAudioPart(content.data!, mimeType: content.mediaType),
          ],
        ),
    ];