extractTextFromResponse static method

String extractTextFromResponse(String jsonStr)

Extract text from a LiteRT-LM JSON response chunk.

Partial / non-JSON chunks pass through verbatim. This is the only shape we are permissive about — any other parse error (TypeError, RangeError, etc.) signals a real contract change with LiteRT-LM and must surface, not be silently swallowed.

Implementation

static String extractTextFromResponse(String jsonStr) {
  final Map<String, dynamic> json;
  try {
    json = jsonDecode(jsonStr) as Map<String, dynamic>;
  } on FormatException {
    return jsonStr;
  }

  final channels = json['channels'] as Map<String, dynamic>?;
  if (channels != null) {
    final thought = channels['thought'] as String?;
    if (thought != null && thought.isNotEmpty) {
      return '<|channel>thought\n$thought<channel|>';
    }
  }

  // The web SDK can represent a text message as a plain string, while the
  // native SDK normally uses the multimodal content-item array. Keep the
  // two supported shapes explicit: an unexpected shape signals a contract
  // change with LiteRT-LM and must not be silently passed into the chat.
  switch (json['content']) {
    case final String text:
      return text;
    case null:
      return jsonStr;
    case final List<Object?> parts:
      final buffer = StringBuffer();
      for (final part in parts) {
        if (part case {'type': 'text', 'text': final String text}) {
          buffer.write(text);
        }
      }
      return buffer.toString();
    case final other:
      throw StateError(
        'LiteRT-LM response content is ${other.runtimeType}, '
        'not a string or a list of parts: $jsonStr',
      );
  }
}