rasa-pro 3.13.8__py3-none-any.whl → 3.14.0a2__py3-none-any.whl

rasa/builder/copilot/copilot.py

@@ -0,0 +1,354 @@
+import asyncio
+import importlib
+import json
+from contextlib import asynccontextmanager
+from typing import Any, Dict, List, Optional
+
+import openai
+import structlog
+from jinja2 import Template
+from typing_extensions import AsyncGenerator
+
+from rasa.builder import config
+from rasa.builder.copilot.constants import (
+    COPILOT_PROMPTS_DIR,
+    COPILOT_PROMPTS_FILE,
+    ROLE_COPILOT,
+    ROLE_COPILOT_INTERNAL,
+    ROLE_SYSTEM,
+    ROLE_USER,
+)
+from rasa.builder.copilot.exceptions import CopilotStreamError
+from rasa.builder.copilot.models import (
+    CopilotContext,
+    ResponseCategory,
+    UsageStatistics,
+)
+from rasa.builder.document_retrieval.inkeep_document_retrieval import (
+    InKeepDocumentRetrieval,
+)
+from rasa.builder.document_retrieval.models import Document
+from rasa.builder.exceptions import (
+    DocumentRetrievalError,
+)
+from rasa.builder.shared.tracker_context import TrackerContext
+from rasa.shared.constants import PACKAGE_NAME
+
+structlogger = structlog.get_logger()
+
+
+class Copilot:
+    def __init__(self) -> None:
+        self._inkeep_document_retrieval = InKeepDocumentRetrieval()
+
+        # Load both prompt templates up-front
+        self._system_message_prompt_template = Template(
+            importlib.resources.read_text(
+                f"{PACKAGE_NAME}.{COPILOT_PROMPTS_DIR}",
+                COPILOT_PROMPTS_FILE,
+            )
+        )
+
+        # The final stream chunk includes usage statistics.
+        self.usage_statistics = UsageStatistics()
+
+    @asynccontextmanager
+    async def _get_client(self) -> AsyncGenerator[openai.AsyncOpenAI, None]:
+        """Create a fresh OpenAI client, yield it, and always close it."""
+        client = openai.AsyncOpenAI(timeout=config.OPENAI_TIMEOUT)
+        try:
+            yield client
+        except Exception as e:
+            structlogger.error("copilot.llm_client_error", error=str(e))
+            raise
+        finally:
+            try:
+                await client.close()
+            except Exception as exc:
+                # Closing should not break request processing, but we log it
+                structlogger.warning(
+                    "copilot.llm_client_close_error",
+                    event_info="Failed to close OpenAI client cleanly.",
+                    error=str(exc),
+                )
+
+    async def search_rasa_documentation(
+        self,
+        context: CopilotContext,
+    ) -> List[Document]:
+        """Search Rasa documentation for relevant information.
+
+        Args:
+            context: The context of the copilot.
+
+        Returns:
+            A list of Document objects. Empty list is returned if the search fails.
+        """
+        try:
+            query = self._create_documentation_search_query(context)
+            return await self._inkeep_document_retrieval.retrieve_documents(query)
+        except DocumentRetrievalError as e:
+            structlogger.error(
+                "copilot.search_rasa_documentation.error",
+                event_info=(
+                    f"Copilot: Searching Rasa documentation for query '{query}' "
+                    f"failed with the following error: {e}. Returning empty list."
+                ),
+                query=query,
+                error=str(e),
+            )
+            return []
+
+    async def generate_response(
+        self,
+        context: CopilotContext,
+    ) -> tuple[AsyncGenerator[str, None], list[Document], str]:
+        """Generate a response from the copilot.
+
+        This method performs document retrieval and response generation as a single
+        atomic operation. The returned documents are the supporting evidence used
+        to generate the response, ensuring consistency between the response content
+        and its sources.
+
+        Args:
+            context: The context of the copilot.
+
+        Returns:
+            A tuple containing the async response stream, relevant documents used
+            as supporting evidence for the generated response, and the prompt used.
+
+        Raises:
+            CopilotStreamError: If the stream fails.
+            Exception: If an unexpected error occurs.
+        """
+        relevant_documents = await self.search_rasa_documentation(context)
+        system_message = await self._create_system_message(context, relevant_documents)
+        chat_history = self._create_chat_history_messages(context)
+        messages = [system_message, *chat_history]
+
+        return (
+            self._stream_response(messages),
+            relevant_documents,
+            system_message.get("content", ""),
+        )
+
+    async def _stream_response(
+        self, messages: List[Dict[str, Any]]
+    ) -> AsyncGenerator[str, None]:
+        """Stream markdown chunks one by one."""
+        self.usage_statistics.reset()
+
+        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},
+                )
+                async for chunk in stream:
+                    # The final chunk, which contains the usage statistics,
+                    # arrives with an empty `choices` list.
+                    if not chunk.choices:
+                        self.usage_statistics.update_from_stream_chunk(chunk)
+                        # Nothing to yield – continue to the next chunk.
+                        continue
+
+                    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(
+                "Failed to stream response from OpenAI API."
+            ) from e
+        except asyncio.TimeoutError as e:
+            structlogger.exception(
+                "copilot.stream_response.timeout_error", error=str(e)
+            )
+            raise CopilotStreamError("Request to OpenAI API timed out.") from e
+        except Exception as e:
+            structlogger.exception(
+                "copilot.stream_response.unexpected_error", error=str(e)
+            )
+            raise
+
+    async def _create_system_message(
+        self,
+        context: CopilotContext,
+        relevant_documents: List[Document],
+    ) -> Dict[str, Any]:
+        """Render the correct Jinja template based on desired output_type."""
+        # Format relevant documentation
+        documents = [doc.model_dump() for doc in relevant_documents]
+
+        # Format conversation history
+        conversation = self._format_conversation_history(context.tracker_context)
+
+        # Format current state
+        current_state = self._format_current_state(context.tracker_context)
+
+        # Render template
+        rendered_prompt = self._system_message_prompt_template.render(
+            current_conversation=conversation,
+            current_state=current_state,
+            assistant_logs=context.assistant_logs,
+            assistant_files=context.assistant_files,
+            documentation_results=documents,
+        )
+        return {"role": ROLE_SYSTEM, "content": rendered_prompt}
+
+    def _create_chat_history_messages(
+        self,
+        context: CopilotContext,
+    ) -> List[Dict[str, Any]]:
+        """Create the chat history messages for the copilot.
+
+        Filter out messages with response_category of GUARDRAILS_POLICY_VIOLATION.
+        This will filter out all the user messages that were flagged by guardrails, but
+        also the copilot messages that were produced by guardrails.
+        """
+        # Filter out messages with response_category of GUARDRAILS_POLICY_VIOLATION.
+        # This will filter out all the user messages flagged by guardrails, but also the
+        # copilot messages that were produced.
+        return [
+            message.to_openai_format()
+            for message in context.copilot_chat_history
+            if message.response_category != ResponseCategory.GUARDRAILS_POLICY_VIOLATION
+        ]
+
+    @staticmethod
+    def _create_documentation_search_query(context: CopilotContext) -> str:
+        """Format chat messages between user and copilot for documentation search."""
+        result = ""
+        role_to_prefix = {
+            ROLE_USER: "User",
+            ROLE_COPILOT: "Assistant",
+            ROLE_COPILOT_INTERNAL: "Copilot Internal Request",
+        }
+        for message in context.copilot_chat_history:
+            prefix = role_to_prefix[message.role]
+            text = message.get_text_content().strip()
+            if text:
+                result += f"{prefix}: {text}\n"
+            log_content = message.get_log_content().strip()
+            if log_content:
+                result += f"{prefix}: {log_content}\n"
+
+        return result
+
+    @staticmethod
+    def _format_documents(results: List[Document]) -> Optional[str]:
+        """Format documentation search results as JSON dump to be used in the prompt."""
+        if not results:
+            return None
+
+        formatted_results = {
+            "sources": [
+                {
+                    # Start the reference from 1, not 0.
+                    "idx": idx + 1,
+                    "title": result.title,
+                    "url": result.url,
+                    "content": result.content,
+                }
+                for idx, result in enumerate(results)
+            ]
+        }
+        return json.dumps(formatted_results, ensure_ascii=False, indent=2)
+
+    @staticmethod
+    def _format_conversation_history(tracker_context: Optional[TrackerContext]) -> str:
+        """Format conversation history from TrackerContext using nested turn structure.
+
+        Args:
+            tracker_context: The TrackerContext containing conversation data.
+
+        Returns:
+            A JSON string with turns containing user_input, assistant_response,
+            and context.
+
+        Example:
+            ```json
+            {
+                "conversation_history": [
+                    {
+                        "turn_id": 1,
+                        "USER": {
+                            "text": "I want to transfer money",
+                            "predicted_commands": ["start flow", "set slot", ...]
+                        },
+                        "BOT": [
+                            {"text": "How much would you like to transfer?"}
+                        ],
+                        "other_tracker_events": [
+                            {
+                                "event": "action_executed",
+                                "data": {"action_name": "action_ask_amount"}
+                            },
+                            {
+                                "event": "slot_set",
+                                "data": {
+                                    "slot_name": "amount_of_money",
+                                    "slot_value": 100,
+                                },
+                            }
+                        ]
+                    }
+                ]
+            }
+            ```
+        """
+        conversation_history: Dict[str, Any] = {
+            "conversation_history": [],
+        }
+
+        if not tracker_context or not tracker_context.conversation_turns:
+            return json.dumps(conversation_history, ensure_ascii=False, indent=2)
+
+        conversation_turns: List[Dict[str, Any]] = []
+        user_prefix = "USER"
+        assistant_prefix = "BOT"
+
+        for turn_idx, turn in enumerate(tracker_context.conversation_turns, 1):
+            turn_data: Dict[str, Any] = {"turn_id": turn_idx}
+
+            # Add user if present
+            if turn.user_message:
+                turn_data[user_prefix] = {
+                    "text": turn.user_message.text,
+                    "predicted_commands": turn.user_message.predicted_commands,
+                }
+
+            # Add assistant messages if present
+            if turn.assistant_messages:
+                turn_data[assistant_prefix] = [
+                    {"text": assistant_message.text}
+                    for assistant_message in turn.assistant_messages
+                ]
+
+            # Add other tracker events
+            if turn.context_events:
+                other_events = [event.model_dump() for event in turn.context_events]
+                turn_data["other_tracker_events"] = other_events
+
+            conversation_turns.append(turn_data)
+
+        conversation_history["conversation_history"] = conversation_turns
+        return json.dumps(conversation_history, ensure_ascii=False, indent=2)
+
+    @staticmethod
+    def _format_current_state(tracker_context: Optional[TrackerContext]) -> str:
+        """Format current state from TrackerContext for LLM consumption.
+
+        Args:
+            tracker_context: The TrackerContext containing current state data.
+
+        Returns:
+            A JSON string containing the current state information.
+        """
+        if not tracker_context or not tracker_context.current_state:
+            return json.dumps({}, ensure_ascii=False, indent=2)
+        current_state = tracker_context.current_state.model_dump()
+        return json.dumps(current_state, ensure_ascii=False, indent=2)