rasa-pro 3.14.1__py3-none-any.whl → 3.15.0a1__py3-none-any.whl

rasa/builder/config.py

@@ -13,6 +13,10 @@ OPENAI_VECTOR_STORE_ID = os.getenv(
 )
 OPENAI_MAX_VECTOR_RESULTS = int(os.getenv("OPENAI_MAX_VECTOR_RESULTS", "10"))
 OPENAI_TIMEOUT = int(os.getenv("OPENAI_TIMEOUT", "30"))
+# OpenAI Token Pricing Configuration (per 1,000 tokens)
+COPILOT_INPUT_TOKEN_PRICE = float(os.getenv("COPILOT_INPUT_TOKEN_PRICE", "0.002"))
+COPILOT_OUTPUT_TOKEN_PRICE = float(os.getenv("COPILOT_OUTPUT_TOKEN_PRICE", "0.0005"))
+COPILOT_CACHED_TOKEN_PRICE = float(os.getenv("COPILOT_CACHED_TOKEN_PRICE", "0.002"))
 
 # Server Configuration
 BUILDER_SERVER_HOST = os.getenv("SERVER_HOST", "0.0.0.0")

rasa/builder/copilot/copilot.py

@@ -42,6 +42,7 @@ from rasa.builder.exceptions import (
     DocumentRetrievalError,
 )
 from rasa.builder.shared.tracker_context import TrackerContext
+from rasa.builder.telemetry.copilot_langfuse_telemetry import CopilotLangfuseTelemetry
 from rasa.shared.constants import PACKAGE_NAME
 
 structlogger = structlog.get_logger()
@@ -72,7 +73,11 @@ class Copilot:
         )
 
         # The final stream chunk includes usage statistics.
-        self.usage_statistics = UsageStatistics()
+        self.usage_statistics = UsageStatistics(
+            input_token_price=config.COPILOT_INPUT_TOKEN_PRICE,
+            output_token_price=config.COPILOT_OUTPUT_TOKEN_PRICE,
+            cached_token_price=config.COPILOT_CACHED_TOKEN_PRICE,
+        )
 
     @asynccontextmanager
     async def _get_client(self) -> AsyncGenerator[openai.AsyncOpenAI, None]:
@@ -94,6 +99,16 @@ class Copilot:
                     error=str(exc),
                 )
 
+    @property
+    def llm_config(self) -> Dict[str, Any]:
+        """The LLM config used to generate the response."""
+        return {
+            "model": config.OPENAI_MODEL,
+            "temperature": config.OPENAI_TEMPERATURE,
+            "stream": True,
+            "stream_options": {"include_usage": True},
+        }
+
     async def search_rasa_documentation(
         self,
         context: CopilotContext,
@@ -108,7 +123,9 @@ class Copilot:
         """
         try:
             query = self._create_documentation_search_query(context)
-            return await self._inkeep_document_retrieval.retrieve_documents(query)
+            documents = await self._inkeep_document_retrieval.retrieve_documents(query)
+            # TODO: Log documentation retrieval to Langfuse
+            return documents
         except DocumentRetrievalError as e:
             structlogger.error(
                 "copilot.search_rasa_documentation.error",
@@ -145,11 +162,12 @@ class Copilot:
             Exception: If an unexpected error occurs.
         """
         relevant_documents = await self.search_rasa_documentation(context)
-        messages = await self._build_messages(context, relevant_documents)
         tracker_event_attachments = self._extract_tracker_event_attachments(
             context.copilot_chat_history[-1]
         )
+        messages = await self._build_messages(context, relevant_documents)
 
+        # TODO: Delete this after Langfuse is implemented
         support_evidence = CopilotGenerationContext(
             relevant_documents=relevant_documents,
             system_message=messages[0],
@@ -163,6 +181,7 @@ class Copilot:
             support_evidence,
         )
 
+    @CopilotLangfuseTelemetry.trace_copilot_streaming_generation
     async def _stream_response(
         self, messages: List[Dict[str, Any]]
     ) -> AsyncGenerator[str, None]:
@@ -172,13 +191,10 @@ class Copilot:
         try:
             async with self._get_client() as client:
                 stream = await client.chat.completions.create(
-                    model=config.OPENAI_MODEL,
-                    messages=messages,  # type: ignore
-                    temperature=config.OPENAI_TEMPERATURE,
-                    stream=True,
-                    stream_options={"include_usage": True},
+                    messages=messages,
+                    **self.llm_config,
                 )
-                async for chunk in stream:
+                async for chunk in stream:  # type: ignore[attr-defined]
                     # The final chunk, which contains the usage statistics,
                     # arrives with an empty `choices` list.
                     if not chunk.choices:
@@ -189,6 +205,7 @@ class Copilot:
                     delta = chunk.choices[0].delta
                     if delta and delta.content:
                         yield delta.content
+
         except openai.OpenAIError as e:
             structlogger.exception("copilot.stream_response.api_error", error=str(e))
             raise CopilotStreamError(
@@ -559,4 +576,6 @@ class Copilot:
         """Extract the tracker event attachments from the message."""
         if not isinstance(message, UserChatMessage):
             return []
+        # TODO: Log tracker event attachments to Langfuse only in the case of the
+        #       User chat message.
         return message.get_content_blocks_by_type(EventContent)

rasa/builder/copilot/models.py

@@ -3,6 +3,7 @@ from enum import Enum
 from typing import Any, Dict, List, Literal, Optional, Type, TypeVar, Union
 
 import structlog
+from openai.types.chat import ChatCompletion
 from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
 from pydantic import (
     BaseModel,
@@ -612,16 +613,171 @@ class TrainingErrorLog(CopilotOutput):
 
 
 class UsageStatistics(BaseModel):
-    prompt_tokens: Optional[int] = None
-    completion_tokens: Optional[int] = None
-    total_tokens: Optional[int] = None
-    model: Optional[str] = None
+    """Usage statistics for a copilot generation."""
+
+    # Token usage statistics
+    prompt_tokens: Optional[int] = Field(
+        default=None,
+        description=(
+            "Total number of prompt tokens used to generate completion. "
+            "Should include cached prompt tokens."
+        ),
+    )
+    completion_tokens: Optional[int] = Field(
+        default=None,
+        description="Number of generated tokens.",
+    )
+    total_tokens: Optional[int] = Field(
+        default=None,
+        description="Total number of tokens used (input + output).",
+    )
+    cached_prompt_tokens: Optional[int] = Field(
+        default=None,
+        description="Number of cached prompt tokens.",
+    )
+    model: Optional[str] = Field(
+        default=None,
+        description="The model used to generate the response.",
+    )
+
+    # Token prices
+    input_token_price: float = Field(
+        default=0.0,
+        description="Price per 1K input tokens in dollars.",
+    )
+    output_token_price: float = Field(
+        default=0.0,
+        description="Price per 1K output tokens in dollars.",
+    )
+    cached_token_price: float = Field(
+        default=0.0,
+        description="Price per 1K cached tokens in dollars.",
+    )
+
+    @property
+    def non_cached_prompt_tokens(self) -> Optional[int]:
+        """Get the non-cached prompt tokens."""
+        if self.cached_prompt_tokens is not None and self.prompt_tokens is not None:
+            return self.prompt_tokens - self.cached_prompt_tokens
+        return self.prompt_tokens
+
+    @property
+    def non_cached_cost(self) -> Optional[float]:
+        """Calculate the non-cached token cost based on configured pricing."""
+        if self.non_cached_prompt_tokens is None:
+            return None
+        if self.non_cached_prompt_tokens == 0:
+            return 0.0
+
+        return (self.non_cached_prompt_tokens / 1000.0) * self.input_token_price
+
+    @property
+    def cached_cost(self) -> Optional[float]:
+        """Calculate the cached token cost based on configured pricing."""
+        if self.cached_prompt_tokens is None:
+            return None
+        if self.cached_prompt_tokens == 0:
+            return 0.0
+
+        return (self.cached_prompt_tokens / 1000.0) * self.cached_token_price
+
+    @property
+    def input_cost(self) -> Optional[float]:
+        """Calculate the input token cost based on configured pricing.
+
+        The calculation takes into account the cached prompt tokens (if available) too.
+        """
+        # If both non-cached and cached costs are None, there's no input cost
+        if self.non_cached_cost is None and self.cached_cost is None:
+            return None
+
+        # If only non-cached cost is available, return it
+        if self.non_cached_cost is not None and self.cached_cost is None:
+            return self.non_cached_cost
+
+        # If only cached cost is available, return it
+        if self.non_cached_cost is None and self.cached_cost is not None:
+            return self.cached_cost
+
+        # If both are available, return the sum
+        return self.non_cached_cost + self.cached_cost  # type: ignore[operator]
+
+    @property
+    def output_cost(self) -> Optional[float]:
+        """Calculate the output token cost based on configured pricing."""
+        if self.completion_tokens is None:
+            return None
+        if self.completion_tokens == 0:
+            return 0.0
+
+        return (self.completion_tokens / 1000.0) * self.output_token_price
+
+    @property
+    def total_cost(self) -> Optional[float]:
+        """Calculate the total cost based on configured pricing.
+
+        Returns:
+            Total cost in dollars, or None if insufficient data.
+        """
+        if self.input_cost is None or self.output_cost is None:
+            return None
+
+        return self.input_cost + self.output_cost
+
+    def update_token_prices(
+        self,
+        input_token_price: float,
+        output_token_price: float,
+        cached_token_price: float,
+    ) -> None:
+        """Update token prices with provided values.
+
+        Args:
+            input_token_price: Price per 1K input tokens in dollars.
+            output_token_price: Price per 1K output tokens in dollars.
+            cached_token_price: Price per 1K cached tokens in dollars.
+        """
+        self.input_token_price = input_token_price
+        self.output_token_price = output_token_price
+        self.cached_token_price = cached_token_price
+
+    @classmethod
+    def from_chat_completion_response(
+        cls,
+        response: ChatCompletion,
+        input_token_price: float = 0.0,
+        output_token_price: float = 0.0,
+        cached_token_price: float = 0.0,
+    ) -> Optional["UsageStatistics"]:
+        """Create a UsageStatistics object from a ChatCompletionChunk."""
+        if not (usage := getattr(response, "usage", None)):
+            return None
+
+        usage_statistics = cls(
+            input_token_price=input_token_price,
+            output_token_price=output_token_price,
+            cached_token_price=cached_token_price,
+        )
+
+        usage_statistics.prompt_tokens = usage.prompt_tokens
+        usage_statistics.completion_tokens = usage.completion_tokens
+        usage_statistics.total_tokens = usage.total_tokens
+        usage_statistics.model = getattr(response, "model", None)
+
+        # Extract cached tokens if available
+        if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+            usage_statistics.cached_prompt_tokens = getattr(
+                usage.prompt_tokens_details, "cached_tokens", None
+            )
+
+        return usage_statistics
 
     def reset(self) -> None:
         """Reset usage statistics to their default values."""
         self.prompt_tokens = None
         self.completion_tokens = None
         self.total_tokens = None
+        self.cached_prompt_tokens = None
         self.model = None
 
     def update_from_stream_chunk(self, chunk: ChatCompletionChunk) -> None:
@@ -630,14 +786,25 @@ class UsageStatistics(BaseModel):
         Args:
             chunk: The OpenAI stream chunk containing usage statistics.
         """
+        # Reset the usage statistics to their default values
+        self.reset()
+
+        # If the chunk has no usage statistics, return
         if not (usage := getattr(chunk, "usage", None)):
             return
 
+        # Update the usage statistics with the values from the chunk
         self.prompt_tokens = usage.prompt_tokens
         self.completion_tokens = usage.completion_tokens
         self.total_tokens = usage.total_tokens
         self.model = getattr(chunk, "model", None)
 
+        # Extract cached tokens if available
+        if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+            self.cached_prompt_tokens = getattr(
+                usage.prompt_tokens_details, "cached_tokens", None
+            )
+
 
 class SigningContext(BaseModel):
     secret: Optional[str] = Field(None)

rasa/builder/document_retrieval/inkeep_document_retrieval.py

@@ -17,6 +17,7 @@ from rasa.builder.document_retrieval.constants import (
 )
 from rasa.builder.document_retrieval.models import Document
 from rasa.builder.exceptions import DocumentRetrievalError
+from rasa.builder.telemetry.copilot_langfuse_telemetry import CopilotLangfuseTelemetry
 from rasa.shared.utils.io import read_json_file
 
 structlogger = structlog.get_logger()
@@ -88,6 +89,7 @@ class InKeepDocumentRetrieval:
             )
             raise e
 
+    @CopilotLangfuseTelemetry.trace_document_retrieval_generation
     async def _call_inkeep_rag_api(
         self, query: str, temperature: float, timeout: float
     ) -> ChatCompletion:

rasa/builder/download.py

@@ -27,7 +27,7 @@ def _get_pyproject_toml_content(project_id: str) -> str:
         version = "0.1.0"
         description = "Add your description for your Rasa bot here"
         readme = "README.md"
-        dependencies = ["rasa-pro>=3.13"]
+        dependencies = ["rasa-pro>=3.14"]
         requires-python = ">={sys.version_info.major}.{sys.version_info.minor}"
         """
     )

rasa/builder/service.py

@@ -5,6 +5,7 @@ import time
 from http import HTTPStatus
 from typing import Any, Optional
 
+import langfuse
 import structlog
 from sanic import Blueprint, HTTPResponse, response
 from sanic.request import Request
@@ -41,7 +42,6 @@ from rasa.builder.copilot.signing import (
     create_signature_envelope_for_text,
     verify_signature,
 )
-from rasa.builder.copilot.telemetry import CopilotTelemetry
 from rasa.builder.download import create_bot_project_archive
 from rasa.builder.guardrails.constants import (
     BLOCK_SCOPE_PROJECT,
@@ -65,6 +65,7 @@ from rasa.builder.models import (
     ApiErrorResponse,
     AssistantInfo,
     BotData,
+    BotFiles,
     JobCreateResponse,
     JobStatus,
     JobStatusEvent,
@@ -74,6 +75,8 @@ from rasa.builder.models import (
 )
 from rasa.builder.project_generator import ProjectGenerator
 from rasa.builder.shared.tracker_context import TrackerContext
+from rasa.builder.telemetry.copilot_langfuse_telemetry import CopilotLangfuseTelemetry
+from rasa.builder.telemetry.copilot_segment_telemetry import CopilotSegmentTelemetry
 from rasa.core.agent import Agent
 from rasa.core.channels.studio_chat import StudioChatInput
 from rasa.core.exceptions import AgentNotReady
@@ -1020,6 +1023,9 @@ async def download_bot_project(request: Request) -> HTTPResponse:
     schema=str,
 )
 @protected()
+# Disable automatic input/output capture for langfuse tracing
+# This allows manual control over what data is sent to langfuse
+@langfuse.observe(capture_input=False, capture_output=False)
 async def copilot(request: Request) -> None:
     """Handle copilot requests with streaming markdown responses."""
     sse = await request.respond(content_type="text/event-stream")
@@ -1046,9 +1052,12 @@ async def copilot(request: Request) -> None:
             )
             return
 
-        telemetry = CopilotTelemetry(project_id=HELLO_RASA_PROJECT_ID, user_id=user_id)
+        telemetry = CopilotSegmentTelemetry(
+            project_id=HELLO_RASA_PROJECT_ID, user_id=user_id
+        )
         structlogger.debug("builder.copilot.telemetry.request.init")
 
+        # TODO: This can be removed once Langfuse is completed.
         if req.last_message and req.last_message.role == ROLE_USER:
             structlogger.debug("builder.copilot.telemetry.request.user_turn")
             # Offload telemetry logging to a background task
@@ -1088,26 +1097,9 @@ async def copilot(request: Request) -> None:
             return
 
         # 4. Get the necessary context for the copilot
-        tracker = await current_tracker_from_input_channel(request.app, req.session_id)
-        tracker_context = TrackerContext.from_tracker(
-            tracker, max_turns=COPILOT_ASSISTANT_TRACKER_MAX_TURNS
-        )
-        if (
-            tracker_context is not None
-            and llm_service.guardrails_policy_checker is not None
-        ):
-            tracker_context = await llm_service.guardrails_policy_checker.check_assistant_chat_for_policy_violations(  # noqa: E501
-                tracker_context=tracker_context,
-                hello_rasa_user_id=user_id,
-                hello_rasa_project_id=HELLO_RASA_PROJECT_ID,
-                lakera_project_id=LAKERA_ASSISTANT_HISTORY_GUARDRAIL_PROJECT_ID,
-            )
-
-        # Copilot doesn't need to know about the docs and any file that is not a core
-        # assistant file
-        relevant_assistant_files = project_generator.get_bot_files(
-            exclude_docs_directory=True,
-            allowed_file_extensions=["yaml", "yml", "py", "jinja", "jinja2"],
+        tracker_context = await get_tracker_context_for_copilot(request, req, user_id)
+        relevant_assistant_files = get_relevant_assistant_files_for_copilot(
+            project_generator,
         )
         context = CopilotContext(
             tracker_context=tracker_context,
@@ -1162,7 +1154,7 @@ async def copilot(request: Request) -> None:
         async for token in intercepted_stream:
             await sse.send(token.to_sse_event().format())
 
-        # 8. Offload telemetry logging to a background task
+        # 8a. Offload metabase telemetry logging to a background task
         request.app.add_task(
             asyncio.to_thread(
                 telemetry.log_copilot_from_handler,
@@ -1177,9 +1169,27 @@ async def copilot(request: Request) -> None:
                     else None
                 ),
                 tracker_event_attachments=generation_context.tracker_event_attachments,
-                **copilot_client.usage_statistics.model_dump(),
+                model=copilot_client.usage_statistics.model or "N/A",
+                prompt_tokens=copilot_client.usage_statistics.prompt_tokens or 0,
+                cached_prompt_tokens=(
+                    copilot_client.usage_statistics.cached_prompt_tokens or 0
+                ),
+                completion_tokens=(
+                    copilot_client.usage_statistics.completion_tokens or 0
+                ),
+                total_tokens=copilot_client.usage_statistics.total_tokens or 0,
             )
         )
+        # 8b. Setup output trace attributes for Langfuse
+        CopilotLangfuseTelemetry.setup_copilot_endpoint_call_trace_attributes(
+            hello_rasa_project_id=HELLO_RASA_PROJECT_ID or "N/A",
+            chat_id=req.session_id or "N/A",
+            user_id=user_id,
+            request=req,
+            handler=copilot_response_handler,
+            relevant_documents=generation_context.relevant_documents,
+            copilot_context=context,
+        )
 
         # 9. Once the stream is over, extract and send references
         #    if any documents were used
@@ -1365,3 +1375,70 @@ async def _handle_guardrail_violation_and_maybe_block(
 
     await sse.send(message.to_sse_event().format())
     return message
+
+
+@langfuse.observe(capture_input=False, capture_output=False)
+async def get_tracker_context_for_copilot(
+    request: Request,
+    req: CopilotRequest,
+    user_id: str,
+) -> Optional[TrackerContext]:
+    """Check the assistant chat for guardrail policy violations.
+
+    Args:
+        request: The request object.
+        req: The CopilotRequest object.
+        user_id: The user ID.
+
+    Returns:
+        The tracker context if the tracker is available.
+    """
+    tracker = await current_tracker_from_input_channel(request.app, req.session_id)
+    tracker_context = TrackerContext.from_tracker(
+        tracker, max_turns=COPILOT_ASSISTANT_TRACKER_MAX_TURNS
+    )
+    if (
+        tracker_context is not None
+        and llm_service.guardrails_policy_checker is not None
+    ):
+        tracker_context = await llm_service.guardrails_policy_checker.check_assistant_chat_for_policy_violations(  # noqa: E501
+            tracker_context=tracker_context,
+            hello_rasa_user_id=user_id,
+            hello_rasa_project_id=HELLO_RASA_PROJECT_ID,
+            lakera_project_id=LAKERA_ASSISTANT_HISTORY_GUARDRAIL_PROJECT_ID,
+        )
+
+    # Track the retrieved tracker context
+    CopilotLangfuseTelemetry.trace_copilot_tracker_context(
+        tracker_context=tracker_context,
+        max_conversation_turns=COPILOT_ASSISTANT_TRACKER_MAX_TURNS,
+        session_id=req.session_id,
+    )
+
+    return tracker_context
+
+
+@langfuse.observe(capture_input=False, capture_output=False)
+def get_relevant_assistant_files_for_copilot(
+    project_generator: ProjectGenerator,
+) -> BotFiles:
+    """Get the relevant assistant files for the copilot.
+
+    Args:
+        project_generator: The project generator.
+
+    Returns:
+        The relevant assistant files.
+    """
+    # Copilot doesn't need to know about the docs and any file that is not a core
+    # assistant file
+    files = project_generator.get_bot_files(
+        exclude_docs_directory=True,
+        allowed_file_extensions=["yaml", "yml", "py", "jinja", "jinja2"],
+    )
+
+    # Track the retrieved assistant files
+    CopilotLangfuseTelemetry.trace_copilot_relevant_assistant_files(
+        relevant_assistant_files=files,
+    )
+    return files