rasa-pro 3.14.1__py3-none-any.whl → 3.15.0.dev20251027__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,
@@ -343,6 +344,55 @@ ChatMessage = Union[
 ]
 
 
+def create_chat_message_from_dict(message_data: Dict[str, Any]) -> ChatMessage:
+    """Parse a single chat message dictionary into a ChatMessage object.
+
+    This utility function manually parses a chat message dictionary into the
+    appropriate ChatMessage type based on its role field.
+
+    Args:
+        message_data: Dictionary containing chat message data
+
+    Returns:
+        Parsed ChatMessage object
+
+    Raises:
+        ValueError: If an unknown role is encountered
+
+    Example:
+        >>> message_data = {
+        ...     "role": "user",
+        ...     "content": [{"type": "text", "text": "Hello"}]
+        ... }
+        >>> message = parse_chat_message_from_dict(message_data)
+        >>> isinstance(message, UserChatMessage)
+        True
+        >>> message.role
+        'user'
+    """
+    available_roles = [ROLE_USER, ROLE_COPILOT, ROLE_COPILOT_INTERNAL]
+    role = message_data.get("role")
+
+    if role == ROLE_USER:
+        return UserChatMessage(**message_data)
+    elif role == ROLE_COPILOT:
+        return CopilotChatMessage(**message_data)
+    elif role == ROLE_COPILOT_INTERNAL:
+        return InternalCopilotRequestChatMessage(**message_data)
+    else:
+        message = (
+            f"Unknown role '{role}' in chat message. "
+            f"Available roles are: {', '.join(available_roles)}."
+        )
+        structlogger.error(
+            "models.create_chat_message_from_dict.unknown_role",
+            event_info=message,
+            role=role,
+            available_roles=available_roles,
+        )
+        raise ValueError(message)
+
+
 class CopilotContext(BaseModel):
     """Model containing the context used by the copilot to generate a response."""
 
@@ -390,37 +440,40 @@ class CopilotRequest(BaseModel):
 
     @field_validator("copilot_chat_history", mode="before")
     @classmethod
-    def parse_chat_history(cls, v: List[Dict[str, Any]]) -> List[ChatMessage]:
+    def parse_chat_history(
+        cls, v: Union[List[Dict[str, Any]], List[ChatMessage]]
+    ) -> List[ChatMessage]:
         """Manually parse chat history messages based on role field."""
+        # If already parsed ChatMessage objects, return them as-is
+        if (
+            v
+            and isinstance(v, list)
+            and all(isinstance(item, ChatMessage) for item in v)
+        ):
+            return v  # type: ignore[return-value]
+
+        # Check for mixed types (some ChatMessage, some not)
+        if (
+            v
+            and isinstance(v, list)
+            and any(isinstance(item, ChatMessage) for item in v)
+        ):
+            message = (
+                "Mixed types in copilot_chat_history: cannot mix ChatMessage objects"
+                "with other types."
+            )
+            structlog.get_logger().error(
+                "copilot_request.parse_chat_history.mixed_types",
+                event_info=message,
+                chat_history_types=[type(item) for item in v],
+            )
+            raise ValueError(message)
+
+        # Otherwise, parse from dictionaries
         parsed_messages: List[ChatMessage] = []
-        available_roles = [ROLE_USER, ROLE_COPILOT, ROLE_COPILOT_INTERNAL]
         for message_data in v:
-            role = message_data.get("role")
-
-            if role == ROLE_USER:
-                parsed_messages.append(UserChatMessage(**message_data))
-
-            elif role == ROLE_COPILOT:
-                parsed_messages.append(CopilotChatMessage(**message_data))
-
-            elif role == ROLE_COPILOT_INTERNAL:
-                parsed_messages.append(
-                    InternalCopilotRequestChatMessage(**message_data)
-                )
-
-            else:
-                message = (
-                    f"Unknown role '{role}' in chat message. "
-                    f"Available roles are: {', '.join(available_roles)}."
-                )
-                structlogger.error(
-                    "copilot_request.parse_chat_history.unknown_role",
-                    event_info=message,
-                    role=role,
-                    available_roles=available_roles,
-                )
-                raise ValueError(message)
-
+            chat_message = create_chat_message_from_dict(message_data)
+            parsed_messages.append(chat_message)
         return parsed_messages
 
     @property
@@ -612,16 +665,171 @@ 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 +838,25 @@ 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/evaluator/constants.py

@@ -0,0 +1,15 @@
+"""Constants for the evaluator module."""
+
+from pathlib import Path
+
+# Base directory for the rasa package
+BASE_DIR = Path(__file__).parent.parent.parent
+
+# Response classification evaluation results directory
+RESPONSE_CLASSIFICATION_EVALUATION_RESULTS_DIR = (
+    BASE_DIR / "builder" / "evaluator" / "results"
+)
+# Default output filename
+DEFAULT_RESPONSE_CLASSIFICATION_EVALUATION_TEXT_OUTPUT_FILENAME = "run_results.txt"
+# Default YAML output filename
+RESPONSE_CLASSIFICATION_EVALUATION_YAML_OUTPUT_FILENAME = "run_results.yaml"

rasa/builder/evaluator/copilot_executor.py

@@ -0,0 +1,89 @@
+"""Copilot execution utilities for evaluators.
+
+This module provides utilities for running copilot operations in evaluation contexts,
+independent of specific evaluation frameworks like Langfuse.
+"""
+
+from typing import List, Optional
+
+import structlog
+from pydantic import BaseModel
+
+from rasa.builder.config import COPILOT_HANDLER_ROLLING_BUFFER_SIZE
+from rasa.builder.copilot.models import (
+    CopilotContext,
+    CopilotGenerationContext,
+    GeneratedContent,
+    ReferenceSection,
+    ResponseCategory,
+)
+from rasa.builder.llm_service import llm_service
+
+structlogger = structlog.get_logger()
+
+
+class CopilotRunResult(BaseModel):
+    """Result from running the copilot with response handler."""
+
+    complete_response: Optional[str]
+    response_category: Optional[ResponseCategory]
+    reference_section: Optional[ReferenceSection]
+    generation_context: CopilotGenerationContext
+
+
+async def run_copilot_with_response_handler(
+    context: CopilotContext,
+) -> Optional[CopilotRunResult]:
+    """Run the copilot with response handler on the given context.
+
+    This function encapsulates the core copilot execution logic. It handles:
+    - Instantiating the copilot and response handler
+    - Generating a response and extracting the reference section from the given context
+    - Returning structured results
+
+    Args:
+        context: The copilot context to process.
+
+    Returns:
+        CopilotRunResult containing the complete response, category, and generation
+        context, or None if execution fails.
+
+    Raises:
+        Any exceptions from the copilot or response handler execution.
+    """
+    # Instantiate the copilot and response handler
+    copilot = llm_service.instantiate_copilot()
+    copilot_response_handler = llm_service.instantiate_handler(
+        COPILOT_HANDLER_ROLLING_BUFFER_SIZE
+    )
+
+    # Call the copilot to generate a response and handle it with the response
+    # handler
+    (original_stream, generation_context) = await copilot.generate_response(context)
+    intercepted_stream = copilot_response_handler.handle_response(original_stream)
+
+    # Exhaust the stream to get the complete response for evaluation
+    response_chunks: List[str] = []
+    response_category = None
+    async for chunk in intercepted_stream:
+        if not isinstance(chunk, GeneratedContent):
+            continue
+        response_chunks.append(chunk.content)
+        response_category = chunk.response_category
+
+    complete_response = "".join(response_chunks) if response_chunks else None
+
+    # Extract the reference section from the response handler
+    if generation_context.relevant_documents:
+        reference_section = copilot_response_handler.extract_references(
+            generation_context.relevant_documents
+        )
+    else:
+        reference_section = None
+
+    return CopilotRunResult(
+        complete_response=complete_response,
+        response_category=response_category,
+        reference_section=reference_section,
+        generation_context=generation_context,
+    )