quraite 0.0.1__py3-none-any.whl

quraite/adapters/bedrock_agents_adapter.py

@@ -0,0 +1,343 @@
+"""
+Bedrock Agents Adapter
+https://docs.aws.amazon.com/bedrock/latest/userguide/trace-events.html
+"""
+
+import asyncio
+import json
+import os
+import uuid
+from typing import Any, Dict, List, Optional, Union
+
+import boto3
+from botocore.exceptions import ClientError
+
+from quraite.adapters.base import BaseAdapter
+from quraite.logger import get_logger
+from quraite.schema.message import (
+    AgentMessage,
+    AssistantMessage,
+    MessageContentText,
+    ToolCall,
+    ToolMessage,
+)
+from quraite.schema.response import AgentInvocationResponse
+
+logger = get_logger(__name__)
+
+
+class BedrockAgentsAdapter(BaseAdapter):
+    """
+    Bedrock Agents adapter wrapper that converts AWS Bedrock agent
+    to a standardized callable interface (invoke) and converts the output to List[AgentMessage].
+
+    This class wraps any Bedrock Agent and provides:
+    - Synchronous invocation via invoke()
+    - Automatic conversion to List[AgentMessage] format
+    """
+
+    def __init__(
+        self,
+        aws_access_key_id: Optional[str] = None,
+        aws_secret_access_key: Optional[str] = None,
+        aws_session_token: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        agent_alias_id: Optional[str] = None,
+        region_name: Optional[str] = None,
+        agent_name: str = "Bedrock Agent",
+    ):
+        """
+        Initialize with Bedrock agent configuration
+
+        Args:
+            aws_access_key_id: AWS access key ID (defaults to AWS_ACCESS_KEY_ID env var)
+            aws_secret_access_key: AWS secret access key (defaults to AWS_SECRET_ACCESS_KEY env var)
+            aws_session_token: AWS session token (defaults to AWS_SESSION_TOKEN env var)
+            agent_id: Bedrock agent ID (defaults to BEDROCK_AGENT_ID env var)
+            agent_alias_id: Bedrock agent alias ID (defaults to BEDROCK_AGENT_ALIAS_ID env var)
+            region_name: AWS region (defaults to AWS_REGION env var)
+            agent_name: Name of the agent for trajectory metadata
+        """
+        logger.debug(
+            "Initializing BedrockAgentsAdapter (agent_name=%s, region=%s)",
+            agent_name,
+            region_name or os.getenv("AWS_REGION"),
+        )
+        self.agent_id = agent_id or os.getenv("BEDROCK_AGENT_ID")
+        self.agent_alias_id = agent_alias_id or os.getenv("BEDROCK_AGENT_ALIAS_ID")
+        self.region_name = region_name or os.getenv("AWS_REGION")
+        self.agent_name = agent_name
+
+        if not all([self.agent_id, self.agent_alias_id, self.region_name]):
+            raise ValueError(
+                "Missing required configuration. Please provide agent_id, agent_alias_id, "
+                "and region_name either as parameters or environment variables."
+            )
+
+        # Initialize Bedrock client
+        self.bedrock_client = boto3.client(
+            region_name=self.region_name,
+            aws_access_key_id=aws_access_key_id or os.getenv("AWS_ACCESS_KEY_ID"),
+            aws_secret_access_key=aws_secret_access_key
+            or os.getenv("AWS_SECRET_ACCESS_KEY"),
+            aws_session_token=aws_session_token or os.getenv("AWS_SESSION_TOKEN"),
+            service_name="bedrock-agent-runtime",
+        )
+        logger.info(
+            "BedrockAgentsAdapter initialized (agent_id=%s, alias=%s, region=%s)",
+            self.agent_id,
+            self.agent_alias_id,
+            self.region_name,
+        )
+
+    def _convert_bedrock_traces_to_messages(
+        self,
+        traces: List[Dict[str, Any]],
+    ) -> List[AgentMessage]:
+        logger.debug("Converting %d Bedrock trace events to messages", len(traces))
+        if not traces:
+            return []
+
+        # TODO: Handle agents with only knowledge base
+        # It has a modelInvocationInput with KNOWLEDGE_BASE_RESPONSE_GENERATION that
+        # has a system prompt. Discuss and decide how to handle this.
+
+        messages = []
+        for trace in traces:
+            # TODO: handle other trace types as well - https://docs.aws.amazon.com/bedrock/latest/userguide/trace-events.html#trace-understand
+            orchestration_trace = trace.get("trace", {}).get("orchestrationTrace", {})
+            if not orchestration_trace:
+                continue
+
+            if "modelInvocationOutput" in orchestration_trace:
+                model_invocation_output = orchestration_trace["modelInvocationOutput"]
+
+                raw_response_content = model_invocation_output.get(
+                    "rawResponse", {}
+                ).get("content", "")
+                if not raw_response_content:
+                    continue
+
+                try:
+                    parsed_content = json.loads(raw_response_content)
+                    contents = parsed_content.get("content", [])
+                except (json.JSONDecodeError, KeyError, ValueError) as e:
+                    logger.exception("Error parsing Bedrock raw response content")
+
+                if not contents:
+                    continue
+
+                text_content = []
+                tool_calls = []
+                for content in contents:
+                    if content.get("type") == "text":
+                        text_content.append(
+                            MessageContentText(
+                                type="text", text=content.get("text", "")
+                            )
+                        )
+                    # TODO: Revsist this later. Ideally should use this but the tool_call_id does not come in the invocationInput
+                    # so for now using trace_id as the tool_call_id
+                    # elif content.get("type") == "tool_use":
+                    #     tool_calls.append(
+                    #         ToolCall(
+                    #             id=content.get("id", ""),
+                    #             name=content.get("name", ""),
+                    #             arguments=content.get("input", {}),
+                    #         )
+                    #     )
+
+                messages.append(
+                    AssistantMessage(
+                        content=text_content if text_content else None,
+                        tool_calls=tool_calls if tool_calls else None,
+                    )
+                )
+            elif "invocationInput" in orchestration_trace:
+                invocation_input = orchestration_trace["invocationInput"]
+                if invocation_input.get("invocationType") == "KNOWLEDGE_BASE":
+                    kb_input = invocation_input.get("knowledgeBaseLookupInput", {})
+                    tool_call = ToolCall(
+                        id=invocation_input.get("traceId", ""),
+                        name="knowledge_base_lookup",
+                        arguments={
+                            "text": kb_input.get("text", ""),
+                            "knowledgeBaseId": kb_input.get("knowledgeBaseId", ""),
+                        },
+                    )
+
+                    if messages and isinstance(messages[-1], AssistantMessage):
+                        if messages[-1].tool_calls:
+                            messages[-1].tool_calls.append(tool_call)
+                        else:
+                            messages[-1].tool_calls = [tool_call]
+                    else:
+                        messages.append(AssistantMessage(tool_calls=[tool_call]))
+
+                elif invocation_input.get("invocationType") == "ACTION_GROUP":
+                    action_group_input = invocation_input.get(
+                        "actionGroupInvocationInput", {}
+                    )
+                    tool_call = ToolCall(
+                        id=invocation_input.get("traceId", ""),
+                        name=f"{action_group_input.get('actionGroupName', '')}/{action_group_input.get('function', '')}",
+                        arguments={
+                            p["name"]: p["value"]
+                            for p in action_group_input.get("parameters", [])
+                        },
+                    )
+                    messages.append(AssistantMessage(tool_calls=[tool_call]))
+            elif "observation" in orchestration_trace:
+                observation = orchestration_trace["observation"]
+                if observation.get("type") == "KNOWLEDGE_BASE":
+                    kb_output = observation.get("knowledgeBaseLookupOutput", {})
+                    tool_result = json.dumps(kb_output.get("retrievedReferences", []))
+                    messages.append(
+                        ToolMessage(
+                            tool_name="knowledge_base_lookup",
+                            tool_call_id=observation.get("traceId", ""),
+                            content=[
+                                MessageContentText(type="text", text=str(tool_result))
+                            ],
+                        )
+                    )
+                elif observation.get("type") == "ACTION_GROUP":
+                    action_group_output = observation.get(
+                        "actionGroupInvocationOutput", {}
+                    )
+
+                    try:
+                        tool_result = json.loads(action_group_output.get("text", ""))
+                    except (json.JSONDecodeError, KeyError, ValueError):
+                        tool_result = action_group_output.get("text", "")
+                    messages.append(
+                        ToolMessage(
+                            tool_name="action_group_invocation",
+                            tool_call_id=observation.get("traceId", ""),
+                            content=[
+                                MessageContentText(type="text", text=str(tool_result))
+                            ],
+                        )
+                    )
+
+        logger.info("Converted Bedrock traces into %d messages", len(messages))
+        return messages
+
+    def _prepare_input(self, input_data: List[AgentMessage]) -> str:
+        """Extract user message from List[AgentMessage]."""
+        logger.debug("Preparing Bedrock input from %d messages", len(input_data))
+        last_user_message = input_data[-1]
+        if last_user_message.role != "user":
+            logger.error("Last message is not from user")
+            return ""
+        # Check if content list is not empty and has text
+        if not last_user_message.content:
+            logger.error("User message has no content")
+            raise ValueError("User message has no content")
+        # Find the first text content item
+        for content_item in last_user_message.content:
+            if content_item.type == "text" and content_item.text:
+                logger.debug(
+                    "Prepared Bedrock input (text_length=%d)", len(content_item.text)
+                )
+                return content_item.text
+        raise ValueError("No text content found in user message")
+
+    def _run_agent(self, session_id: str, prompt: str) -> List[Dict]:
+        """
+        Run the Bedrock agent and collect response and traces.
+
+        Args:
+            session_id: Unique session identifier
+            prompt: Input prompt for the agent
+
+        Returns:
+            List of traces
+        """
+        try:
+            agent_answer = ""
+            logger.debug(
+                "Invoking Bedrock agent (session_id=%s, prompt_length=%d)",
+                session_id,
+                len(prompt),
+            )
+            response = self.bedrock_client.invoke_agent(
+                agentId=self.agent_id,
+                agentAliasId=self.agent_alias_id,
+                sessionId=session_id,
+                inputText=prompt,
+                enableTrace=True,
+            )
+
+            stream = response["completion"]
+            traces = []
+
+            for event in stream:
+                if "chunk" in event:
+                    data = event["chunk"]["bytes"]
+                    agent_answer = data.decode("utf8")
+                event_trace = event.get("trace")
+
+                if event_trace:
+                    traces.append(event_trace)
+
+        except ClientError:
+            logger.exception("Error invoking Bedrock agent via Bedrock runtime")
+            return "", []
+        logger.info(
+            "Bedrock agent invocation succeeded (session_id=%s, trace_events=%d)",
+            session_id,
+            len(traces),
+        )
+        return agent_answer, traces
+
+    async def ainvoke(
+        self,
+        input: List[AgentMessage],
+        session_id: Union[str, None],
+    ) -> AgentInvocationResponse:
+        """Asynchronous invocation method - invokes the Bedrock agent and converts to List[AgentMessage].
+
+        Args:
+            input: List of AgentMessage objects
+            session_id: Unique session identifier
+
+        Returns:
+            AgentInvocationResponse - response containing agent trace, trajectory, and final response.
+        """
+        logger.info(
+            "Bedrock ainvoke called (session_id=%s, input_messages=%d)",
+            session_id,
+            len(input),
+        )
+        agent_input = self._prepare_input(input)
+        session_id = session_id or str(uuid.uuid4())
+
+        try:
+            # Run the synchronous _run_agent in a thread pool to avoid blocking
+            _, traces = await asyncio.to_thread(
+                self._run_agent, session_id, agent_input
+            )
+            logger.debug(
+                "Bedrock agent run returned %d trace events for session_id=%s",
+                len(traces),
+                session_id,
+            )
+        except (ClientError, ValueError, KeyError, json.JSONDecodeError):
+            logger.exception("Error invoking Bedrock agent")
+            return AgentInvocationResponse()
+
+        try:
+            agent_trajectory = self._convert_bedrock_traces_to_messages(traces)
+            logger.info(
+                "Bedrock agent produced %d trajectory messages",
+                len(agent_trajectory),
+            )
+
+            return AgentInvocationResponse(
+                agent_trajectory=agent_trajectory,
+            )
+
+        except (ClientError, ValueError, KeyError, json.JSONDecodeError):
+            logger.exception("Error converting Bedrock traces to messages")
+            return AgentInvocationResponse()

quraite/adapters/flowise_adapter.py

@@ -0,0 +1,275 @@
+import asyncio
+import json
+from typing import Any, Dict, List, Optional, Union
+
+import aiohttp
+
+from quraite.adapters.base import BaseAdapter
+from quraite.logger import get_logger
+from quraite.schema.message import (
+    AgentMessage,
+    AssistantMessage,
+    MessageContentText,
+    ToolCall,
+    ToolMessage,
+)
+from quraite.schema.response import AgentInvocationResponse
+
+logger = get_logger(__name__)
+
+
+class FlowiseAdapter(BaseAdapter):
+    def __init__(
+        self, api_url: str, headers: Optional[Dict[str, str]] = None, timeout: int = 60
+    ):
+        self.api_url = api_url
+        self.headers = headers or {}
+        self.timeout = timeout
+
+        if "Content-Type" not in self.headers:
+            self.headers["Content-Type"] = "application/json"
+
+        self.headers["Content-Type"] = "application/json"
+        logger.info(
+            "FlowiseAdapter initialized (api_url=%s, timeout=%s)", self.api_url, timeout
+        )
+
+    AGENT_NODE_NAME = "agentAgentflow"
+
+    def _convert_api_output_to_messages(
+        self,
+        response: Dict[str, Any],
+        request_messages: Optional[List[AgentMessage]] = None,
+    ) -> List[AgentMessage]:
+        logger.debug(
+            "Converting Flowise response to messages (has_agent_node=%s)",
+            bool(response.get("agentFlowExecutedData")),
+        )
+
+        def _append_text(content_obj: Any, bucket: List[MessageContentText]) -> None:
+            if isinstance(content_obj, str):
+                if content_obj:
+                    bucket.append(MessageContentText(type="text", text=content_obj))
+            elif isinstance(content_obj, list):
+                for item in content_obj:
+                    _append_text(item, bucket)
+            elif isinstance(content_obj, dict):
+                text_value = content_obj.get("text")
+                if text_value:
+                    bucket.append(MessageContentText(type="text", text=text_value))
+
+        agent_node = self._find_agentflow_node(response)
+        if agent_node:
+            converted: List[AgentMessage] = []
+            messages_section = agent_node.get("input")
+            raw_messages = []
+            if isinstance(messages_section, dict):
+                potential_messages = messages_section.get("messages")
+                if isinstance(potential_messages, list):
+                    raw_messages = potential_messages
+
+            latest_turn: List[Dict[str, Any]] = []
+            for raw in reversed(raw_messages):
+                if not isinstance(raw, dict):
+                    continue
+                if raw.get("role") == "user":
+                    break
+                latest_turn.append(raw)
+            latest_turn.reverse()
+
+            for raw in latest_turn:
+                role = raw.get("role")
+                if role not in {"assistant", "system", "tool"}:
+                    continue
+
+                if role == "assistant":
+                    text_content: List[MessageContentText] = []
+                    _append_text(raw.get("content"), text_content)
+                    tool_calls_list: List[ToolCall] = []
+                    tool_calls = raw.get("tool_calls")
+                    if isinstance(tool_calls, list):
+                        for call in tool_calls:
+                            if not isinstance(call, dict):
+                                continue
+                            arguments = call.get("args")
+                            if not isinstance(arguments, dict):
+                                arguments = {"value": arguments}
+                            tool_calls_list.append(
+                                ToolCall(
+                                    id=call.get("id"),
+                                    name=call.get("name"),
+                                    arguments=arguments,
+                                )
+                            )
+
+                    converted.append(
+                        AssistantMessage(
+                            content=text_content if text_content else None,
+                            tool_calls=tool_calls_list if tool_calls_list else None,
+                        )
+                    )
+
+                elif role == "tool":
+                    payload = raw.get("content")
+                    tool_result = (
+                        payload if isinstance(payload, dict) else {"output": payload}
+                    )
+                    converted.append(
+                        ToolMessage(
+                            tool_name=raw.get("name"),
+                            tool_call_id=raw.get("tool_call_id"),
+                            content=[
+                                MessageContentText(
+                                    type="text", text=json.dumps(tool_result)
+                                )
+                            ],
+                        )
+                    )
+
+            output_section = agent_node.get("output")
+            if isinstance(output_section, dict):
+                final_segments: List[MessageContentText] = []
+                _append_text(output_section.get("content"), final_segments)
+
+                if final_segments:
+                    converted.append(AssistantMessage(content=final_segments))
+
+            if converted:
+                return converted
+
+        if request_messages:
+            logger.debug("Falling back to request messages; no agent output found")
+            return list(request_messages)
+
+        return []
+
+    def _find_agentflow_node(
+        self, response: Dict[str, Any]
+    ) -> Optional[Dict[str, Any]]:
+        executed_data = response.get("agentFlowExecutedData")
+        if not isinstance(executed_data, list):
+            return None
+
+        for entry in executed_data:
+            node_data = entry.get("data")
+            if not isinstance(node_data, dict):
+                continue
+
+            if node_data.get("name") == self.AGENT_NODE_NAME:
+                return node_data
+
+        return None
+
+    def _prepare_input(self, input: List[AgentMessage]) -> str:
+        logger.debug("Preparing Flowise input from %d messages", len(input))
+        if not input or input[-1].role != "user":
+            logger.error("Flowise input missing user message")
+            raise ValueError("No user message found in the input")
+
+        last_user_message = input[-1]
+        if not last_user_message.content:
+            logger.error("Flowise input user message missing content")
+            raise ValueError("User message has no content")
+
+        text_content = None
+        for content_item in last_user_message.content:
+            if content_item.type == "text" and content_item.text:
+                text_content = content_item.text
+                break
+
+        if not text_content:
+            logger.error("Flowise input missing text content")
+            raise ValueError("No text content found in user message")
+
+        logger.debug("Prepared Flowise input (text_length=%d)", len(text_content))
+        return text_content
+
+    async def _aapi_call(
+        self,
+        question: str,
+        history: Optional[List[Dict[str, Any]]] = None,
+        override_config: Optional[Dict[str, Any]] = None,
+        form: Optional[Dict[str, Any]] = None,
+        human_input: Optional[str] = None,
+        **kwargs,
+    ) -> Dict[str, Any]:
+        logger.debug(
+            "Calling Flowise API (history=%s, form=%s, human_input=%s)",
+            bool(history),
+            bool(form),
+            bool(human_input),
+        )
+        payload = {
+            "question": question,
+            "overrideConfig": override_config or {},
+            **kwargs,
+        }
+
+        if history is not None:
+            payload["history"] = history
+
+        if form is not None:
+            payload["form"] = form
+
+        if human_input is not None:
+            payload["humanInput"] = human_input
+
+        async with aiohttp.ClientSession() as session:
+            try:
+                async with session.post(
+                    self.api_url,
+                    headers=self.headers,
+                    json=payload,
+                    timeout=aiohttp.ClientTimeout(total=self.timeout),
+                ) as response:
+                    response.raise_for_status()
+                    logger.info(
+                        "Flowise API call succeeded (status=%s)", response.status
+                    )
+                    return await response.json()
+
+            except (aiohttp.ClientError, asyncio.TimeoutError) as e:
+                logger.exception("Flowise API request failed")
+                raise aiohttp.ClientError(f"Async API request failed: {str(e)}") from e
+
+            except json.JSONDecodeError as e:
+                logger.exception("Flowise API response decoding failed")
+                raise ValueError(f"Failed to decode JSON response: {e}") from e
+
+    async def ainvoke(
+        self,
+        input: List[AgentMessage],
+        session_id: Union[str, None],
+    ) -> AgentInvocationResponse:
+        """Asynchronous invocation method - invokes the Flowise agent and converts to List[AgentMessage]."""
+        logger.info(
+            "Flowise ainvoke called (session_id=%s, input_messages=%d)",
+            session_id,
+            len(input),
+        )
+        agent_input = self._prepare_input(input)
+
+        try:
+            agent_output = await self._aapi_call(
+                question=agent_input,
+                override_config={"sessionId": session_id} if session_id else None,
+            )
+            logger.debug(
+                "Flowise API returned payload with keys: %s", list(agent_output.keys())
+            )
+        except Exception as e:
+            logger.exception("Flowise ainvoke failed during API call")
+            raise RuntimeError(f"Error calling Flowise endpoint: {e}") from e
+
+        try:
+            agent_trajectory = self._convert_api_output_to_messages(agent_output, input)
+            logger.info(
+                "Flowise conversion produced %d trajectory messages",
+                len(agent_trajectory),
+            )
+            return AgentInvocationResponse(
+                agent_trajectory=agent_trajectory,
+            )
+        except Exception as e:
+            logger.exception("Flowise ainvoke failed while processing response")
+            raise RuntimeError(f"Error processing Flowise response: {e}") from e