Merge branch 'main' into feat/vici-dial

api/services/configuration/registry.py

@@ -584,6 +584,20 @@ class SarvamLLMConfiguration(BaseLLMConfiguration):
 
 
 OPENAI_REALTIME_MODELS = ["gpt-realtime-2"]
+# ISO 639-1 codes accepted by the Realtime API's input_audio_transcription.
+# Not exhaustive — the field allows custom input.
+OPENAI_REALTIME_LANGUAGES = [
+    "en",
+    "es",
+    "pt",
+    "fr",
+    "de",
+    "it",
+    "hi",
+    "ja",
+    "ko",
+    "zh",
+]
 OPENAI_REALTIME_VOICES = [
     "alloy",
     "ash",
@@ -618,6 +632,17 @@ class OpenAIRealtimeLLMConfiguration(BaseLLMConfiguration):
             "allow_custom_input": True,
         },
     )
+    language: str | None = Field(
+        default=None,
+        description=(
+            "ISO 639-1 language code for input audio transcription (e.g. 'pt', 'es'). "
+            "Improves transcription accuracy and latency. Leave unset to auto-detect."
+        ),
+        json_schema_extra={
+            "examples": OPENAI_REALTIME_LANGUAGES,
+            "allow_custom_input": True,
+        },
+    )
 
 
 GROK_REALTIME_MODELS = ["grok-voice-think-fast-1.0"]

api/services/pipecat/audio_config.py

@@ -52,16 +52,6 @@ class AudioConfig:
             )
             self.pipeline_sample_rate = 16000
 
-        # Log configuration for auditing
-        logger.info(
-            f"AudioConfig initialized: "
-            f"transport_in={self.transport_in_sample_rate}Hz, "
-            f"transport_out={self.transport_out_sample_rate}Hz, "
-            f"vad={self.vad_sample_rate}Hz, "
-            f"pipeline={self.pipeline_sample_rate}Hz, "
-            f"buffer={self.buffer_size_seconds}s"
-        )
-
     @property
     def buffer_size_bytes(self) -> int:
         """Calculate buffer size in bytes based on pipeline sample rate."""

api/services/pipecat/run_pipeline.py

@@ -95,6 +95,9 @@ from pipecat.turns.user_start import (
     MinWordsUserTurnStartStrategy,
     ProvisionalVADUserTurnStartStrategy,
 )
+from pipecat.turns.user_start.transcription_user_turn_start_strategy import (
+    TranscriptionUserTurnStartStrategy,
+)
 from pipecat.turns.user_start.vad_user_turn_start_strategy import (
     VADUserTurnStartStrategy,
 )
@@ -160,9 +163,10 @@ def _create_non_realtime_user_turn_start_strategies(
 
     if turn_start_strategy == "provisional_vad":
         return [
+            TranscriptionUserTurnStartStrategy(),
             ProvisionalVADUserTurnStartStrategy(
                 pause_secs=_resolve_provisional_vad_pause_secs(run_configs)
-            )
+            ),
         ]
 
     if uses_external_turns:
@@ -172,7 +176,7 @@ def _create_non_realtime_user_turn_start_strategies(
         # confirms a real turn.
         return [ExternalUserTurnStartStrategy(enable_interruptions=True)]
 
-    return [VADUserTurnStartStrategy()]
+    return [TranscriptionUserTurnStartStrategy(), VADUserTurnStartStrategy()]
 
 
 def _create_non_realtime_user_turn_stop_strategies(

api/services/pipecat/service_factory.py

@@ -227,7 +227,6 @@ def create_stt_service(
         # Other models than flux
         # Use language from user config, defaulting to "multi" for multilingual support
         language = getattr(user_config.stt, "language", None) or "multi"
-        logger.debug(f"Using DeepGram Model - {user_config.stt.model}")
         return DeepgramSTTService(
             api_key=user_config.stt.api_key,
             settings=DeepgramSTTSettings(
@@ -1009,6 +1008,13 @@ def create_realtime_llm_service(user_config, audio_config: "AudioConfig"):
             SessionProperties,
         )
 
+        # Pin the transcription language when configured. Without it the model
+        # auto-detects per utterance, which misfires on short/noisy telephony
+        # audio (e.g. Portuguese transcribed as English or Chinese).
+        transcription_kwargs = {}
+        if language:
+            transcription_kwargs["language"] = language
+
         return DograhOpenAIRealtimeLLMService(
             api_key=api_key,
             settings=DograhOpenAIRealtimeLLMService.Settings(
@@ -1016,7 +1022,9 @@ def create_realtime_llm_service(user_config, audio_config: "AudioConfig"):
                 session_properties=SessionProperties(
                     audio=AudioConfiguration(
                         input=AudioInput(
-                            transcription=InputAudioTranscription(),
+                            transcription=InputAudioTranscription(
+                                **transcription_kwargs
+                            ),
                         ),
                         output=AudioOutput(
                             voice=voice or "alloy",

api/services/workflow/tools/custom_tool.py

@@ -8,6 +8,7 @@ import httpx
 from loguru import logger
 
 from api.db import db_client
+from api.services.configuration.masking import mask_key
 from api.utils.credential_auth import build_auth_header
 from api.utils.template_renderer import render_template
 
@@ -21,6 +22,19 @@ TYPE_MAP = {
 }
 
 
+def serialize_query_params(arguments: Dict[str, Any]) -> Dict[str, Any]:
+    """JSON-stringify dict/list values so they're safe to pass as query params.
+
+    httpx (and query strings in general) only support primitive param values.
+    Object/array-typed tool arguments must be serialized before going out as
+    GET/DELETE query params, otherwise httpx raises a TypeError.
+    """
+    return {
+        k: json.dumps(v) if isinstance(v, (dict, list)) else v
+        for k, v in arguments.items()
+    }
+
+
 def tool_to_function_schema(tool: Any) -> Dict[str, Any]:
     """Convert a ToolModel to an LLM function schema.
 
@@ -224,7 +238,9 @@ async def execute_http_tool(
     arguments: Dict[str, Any],
     call_context_vars: Optional[Dict[str, Any]] = None,
     gathered_context_vars: Optional[Dict[str, Any]] = None,
+    preset_params: Optional[Dict[str, Any]] = None,
     organization_id: Optional[int] = None,
+    include_request_headers: bool = False,
 ) -> Dict[str, Any]:
     """Execute an HTTP API tool.
 
@@ -233,7 +249,11 @@ async def execute_http_tool(
         arguments: Arguments passed by the LLM (parameter name -> value)
         call_context_vars: Initial context variables available at runtime
         gathered_context_vars: Variables extracted during the conversation
+        preset_params: Pre-resolved preset parameter values. Used by the test
+            endpoint; live calls omit this so configured templates are resolved.
         organization_id: Organization ID for credential lookup
+        include_request_headers: Include a client-safe header preview in the result.
+            Headers supplied by a stored credential are masked.
 
     Returns:
         Result dict with response data or error
@@ -248,7 +268,9 @@ async def execute_http_tool(
     # Get headers from config
     headers = dict(config.get("headers", {}) or {})
 
-    # Add auth header if credential is configured
+    # Add auth header if credential is configured. Keep track of which headers
+    # came from the credential so only those values are masked in test previews.
+    credential_headers: Dict[str, str] = {}
     credential_uuid = config.get("credential_uuid")
     if credential_uuid and organization_id:
         try:
@@ -256,8 +278,8 @@ async def execute_http_tool(
                 credential_uuid, organization_id
             )
             if credential:
-                auth_header = build_auth_header(credential)
-                headers.update(auth_header)
+                credential_headers = build_auth_header(credential)
+                headers.update(credential_headers)
                 logger.debug(f"Applied credential '{credential.name}' to tool request")
             else:
                 logger.warning(
@@ -266,17 +288,31 @@ async def execute_http_tool(
         except Exception as e:
             logger.error(f"Failed to fetch credential for tool '{tool.name}': {e}")
 
+    request_headers: Dict[str, str] = {}
+    if include_request_headers:
+        request_headers = {str(name): str(value) for name, value in headers.items()}
+        for header_name, header_value in credential_headers.items():
+            request_headers[header_name] = mask_key(str(header_value))
+
+    def build_result(result: Dict[str, Any]) -> Dict[str, Any]:
+        if include_request_headers:
+            return {**result, "request_headers": request_headers}
+        return result
+
     # Get timeout
     timeout_ms = config.get("timeout_ms", 5000)
     timeout_seconds = timeout_ms / 1000
 
-    try:
-        preset_arguments = _resolve_preset_parameters(
-            config, call_context_vars, gathered_context_vars
-        )
-    except ValueError as e:
-        logger.error(f"Custom tool '{tool.name}' preset parameter error: {e}")
-        return {"status": "error", "error": str(e)}
+    if preset_params is None:
+        try:
+            preset_arguments = _resolve_preset_parameters(
+                config, call_context_vars, gathered_context_vars
+            )
+        except ValueError as e:
+            logger.error(f"Custom tool '{tool.name}' preset parameter error: {e}")
+            return build_result({"status": "error", "error": str(e)})
+    else:
+        preset_arguments = dict(preset_params)
 
     resolved_arguments = {**(arguments or {}), **preset_arguments}
 
@@ -286,7 +322,7 @@ async def execute_http_tool(
     if method in ("POST", "PUT", "PATCH"):
         body = resolved_arguments
     elif method in ("GET", "DELETE") and resolved_arguments:
-        params = resolved_arguments
+        params = serialize_query_params(resolved_arguments)
 
     logger.info(
         f"Executing custom tool '{tool.name}' ({tool.tool_uuid}): {method} {url}"
@@ -322,23 +358,29 @@ async def execute_http_tool(
             logger.debug(
                 f"Custom tool '{tool.name}' completed with status {response.status_code}"
             )
-            return result
+            return build_result(result)
 
     except httpx.TimeoutException:
         logger.error(f"Custom tool '{tool.name}' timed out after {timeout_seconds}s")
-        return {
-            "status": "error",
-            "error": f"Request timed out after {timeout_seconds} seconds",
-        }
+        return build_result(
+            {
+                "status": "error",
+                "error": f"Request timed out after {timeout_seconds} seconds",
+            }
+        )
     except httpx.RequestError as e:
         logger.error(f"Custom tool '{tool.name}' request failed: {e}")
-        return {
-            "status": "error",
-            "error": f"Request failed: {str(e)}",
-        }
+        return build_result(
+            {
+                "status": "error",
+                "error": f"Request failed: {str(e)}",
+            }
+        )
     except Exception as e:
         logger.error(f"Custom tool '{tool.name}' execution failed: {e}")
-        return {
-            "status": "error",
-            "error": f"Tool execution failed: {str(e)}",
-        }
+        return build_result(
+            {
+                "status": "error",
+                "error": f"Tool execution failed: {str(e)}",
+            }
+        )