* feat: Add ErrorFrame emission to TTS/STT services for pipeline error detection
- Add ErrorFrame emission to all major TTS/STT services during initialization and runtime failures
- Services updated: Cartesia, ElevenLabs, Deepgram, AssemblyAI, Rime, Azure
- ErrorFrame objects emitted with fatal=False for graceful degradation
- Enables on_pipeline_error event handler to detect service failures programmatically
- Add comprehensive pytest test suite to verify ErrorFrame emission
- Fixes issue where services failed gracefully but didn't emit ErrorFrame objects
This allows developers to implement real-time error monitoring and alerting
using the on_pipeline_error event handler introduced in v0.0.90.
* Update STT and TTS services to use consistent error handling pattern
- Improves error handling consistency across all services
* Add changelog entry for STT/TTS error handling improvements
* Linting issues Resolved
* Azure STT ErrorFrames added with consistent patterns
* Cartesia STT and Deepgram STT; additional fixes made
* Removed Fatal Flags across services, removed duplication
* Moving the changelog entry to the correct place.
* Refactoring some classes to use yield instead of push_error directly.
* Fixing ruff format.
---------
Co-authored-by: Filipi Fuchter <filipi87@gmail.com>
Note that for `LLMTextFrame`s, the right behavior is pretty much always `includes_inter_frame_spaces = True`. I decided *not* to go ahead and make that the default for `LLMTextFrame`s, though, simply to not introduce a subtle behavior change for creative/unexpected use-cases that were relying on text in hand-crafted `LLMTextFrame`s being handled a certain way. Ditto for `TTSTextFrame`s.
Also, fix an issue in `NeuphonicTTSService` where it wasn't pushing `TTSTextFrame`s.
Also, fix the broken `SarvamHttpTTSService` example.
Also, add a couple of missing examples.
* Fix Langfuse tracing for GoogleLLMService with universal LLMContext
- Fixed issue where input appeared as null in Langfuse dashboard for GoogleLLMService
- Added fallback to use adapter's get_messages_for_logging() for universal LLMContext
- Ensures proper message format conversion for Google/Gemini services
- Handles system message conversion to system_instruction format
- Also fixes serialization of empty message lists ([] now serializes correctly)
This fix ensures Langfuse tracing works correctly for Google services using
both OpenAILLMContext/GoogleLLMContext and the universal LLMContext.
* Add unit tests for Langfuse tracing with GoogleLLMService
- Test that tracing correctly captures messages with universal LLMContext
- Test that empty message lists are properly serialized
- Test that adapter's get_messages_for_logging is used instead of context method
- All tests verify that input is correctly added to Langfuse spans
* Fix test mocking to patch opentelemetry.trace.get_tracer correctly
The tests were failing in CI because they were trying to patch
'pipecat.utils.tracing.service_decorators.trace' which doesn't exist as
an attribute. The trace module is imported from opentelemetry, so we need
to patch 'opentelemetry.trace.get_tracer' instead.
* Skip tracing tests when opentelemetry is not installed
The tracing dependencies (opentelemetry) are optional in Pipecat and not
installed in the CI environment. Added a skipif marker to skip these tests
when opentelemetry is not available, preventing CI failures while still
allowing the tests to run when tracing dependencies are installed locally.
* Install tracing dependencies in GitHub Actions CI
Instead of skipping the tracing tests, install the 'tracing' extra
(opentelemetry) in the CI environment so the tests can run properly.
Removed the skipif condition from the tests since opentelemetry will
now be available in CI.
* Use the context type to determine which messages to use, fix tool_count and tools (#3032)
---------
Co-authored-by: Mark Backman <mark@daily.co>
I chose to go the somewhat hacky route of adding the `ToolsSchema` support into the `events.SessionProperties` model itself—even though we should never serialize that type when creating events—because the alternative seemed to be to create a new type for `OpenAIRealtimeLLMService` initialization parameters and then we'd have to contend with backward compatibility, which seemed like a bigger headache.
This allows folks to use `MCPClient` alongside the pattern of passing in tools at LLM init time, a pattern supported by speech-to-speech services such as `GeminiLiveLLMService`.
This does a couple of things:
- Makes the `MCPClient` LLM agnostic, setting us up for some upcoming improvements (like making it possible to use with `LLMSwitcher`)
- Makes `GeminiLLMAdapter` more robust, as the schema massaging that was previously only done in `MCPClient` is useful for all tools, not just for MCP-provided ones
