VibecoderMcSwaggins's picture
fix(arch): End-to-end BYOK support for unified architecture
915b009
raw
history blame
22.6 kB
"""
Advanced Orchestrator using Microsoft Agent Framework.
This orchestrator uses the ChatAgent pattern from Microsoft's agent-framework-core
package for multi-agent coordination. It provides richer orchestration capabilities
including specialized agents (Search, Hypothesis, Judge, Report) coordinated by
a manager agent.
Note: Previously named 'orchestrator_magentic.py' - renamed to eliminate confusion
with the 'magentic' PyPI package (which is a different library).
Design Patterns:
- Mediator: Manager agent coordinates between specialized agents
- Strategy: Different agents implement different strategies for their tasks
- Observer: Event stream allows UI to observe progress
"""
import asyncio
from collections.abc import AsyncGenerator
from typing import TYPE_CHECKING, Any, Literal
import structlog
from agent_framework import (
MagenticAgentDeltaEvent,
MagenticAgentMessageEvent,
MagenticBuilder,
MagenticFinalResultEvent,
MagenticOrchestratorMessageEvent,
WorkflowOutputEvent,
)
from src.agents.magentic_agents import (
create_hypothesis_agent,
create_judge_agent,
create_report_agent,
create_search_agent,
)
from src.agents.state import init_magentic_state
from src.clients.base import BaseChatClient
from src.clients.factory import get_chat_client
from src.config.domain import ResearchDomain, get_domain_config
from src.orchestrators.base import OrchestratorProtocol
from src.utils.config import settings
from src.utils.models import AgentEvent
from src.utils.service_loader import get_embedding_service_if_available
if TYPE_CHECKING:
from src.services.embedding_protocol import EmbeddingServiceProtocol
logger = structlog.get_logger()
class AdvancedOrchestrator(OrchestratorProtocol):
"""
Advanced orchestrator using Microsoft Agent Framework ChatAgent pattern.
Each agent has an internal LLM that understands natural language
instructions from the manager and can call tools appropriately.
This orchestrator provides:
- Multi-agent coordination (Search, Hypothesis, Judge, Report)
- Manager agent for workflow orchestration
- Streaming events for real-time UI updates
- Configurable timeouts and round limits
"""
# Estimated seconds per coordination round (for progress UI)
_EST_SECONDS_PER_ROUND: int = 45
def __init__(
self,
max_rounds: int = 5,
chat_client: BaseChatClient | None = None,
provider: str | None = None,
api_key: str | None = None,
domain: ResearchDomain | str | None = None,
timeout_seconds: float | None = None,
) -> None:
"""Initialize the advanced orchestrator.
Args:
max_rounds: Maximum number of coordination rounds.
chat_client: Optional pre-configured chat client.
provider: Optional provider override ("openai", "huggingface").
api_key: Optional API key override.
domain: Research domain for customization.
timeout_seconds: Optional timeout override (defaults to settings).
"""
self._max_rounds = max_rounds
self.domain = domain or ResearchDomain.SEXUAL_HEALTH
self.domain_config = get_domain_config(self.domain)
self._timeout_seconds = timeout_seconds or settings.advanced_timeout
self.logger = logger.bind(orchestrator="advanced")
# Use provided client or create one via factory
self._chat_client = chat_client or get_chat_client(
provider=provider,
api_key=api_key,
)
# Store API key for service initialization
self._api_key = api_key
# Event stream for UI updates
self._events: list[AgentEvent] = []
# Initialize services lazily
self._embedding_service: EmbeddingServiceProtocol | None = None
# Track execution statistics
self.stats = {
"rounds": 0,
"searches": 0,
"hypotheses": 0,
"reports": 0,
"errors": 0,
}
def _init_embedding_service(self) -> "EmbeddingServiceProtocol | None":
"""Initialize embedding service if available."""
return get_embedding_service_if_available(api_key=self._api_key)
def _build_workflow(self) -> Any:
"""Build the workflow with ChatAgent participants."""
# Create agents with internal LLMs
search_agent = create_search_agent(self._chat_client, domain=self.domain)
judge_agent = create_judge_agent(self._chat_client, domain=self.domain)
hypothesis_agent = create_hypothesis_agent(self._chat_client, domain=self.domain)
report_agent = create_report_agent(self._chat_client, domain=self.domain)
# Manager chat client (orchestrates the agents)
manager_client = self._chat_client
return (
MagenticBuilder()
.participants(
searcher=search_agent,
hypothesizer=hypothesis_agent,
judge=judge_agent,
reporter=report_agent,
)
.with_standard_manager(
chat_client=manager_client,
max_round_count=self._max_rounds,
max_stall_count=3,
max_reset_count=2,
)
.build()
)
def _create_task_prompt(self, query: str) -> str:
"""Create the initial task prompt for the manager agent."""
return f"""Research {self.domain_config.report_focus} for: {query}
## CRITICAL RULE
When JudgeAgent says "SUFFICIENT EVIDENCE" or "STOP SEARCHING":
β†’ IMMEDIATELY delegate to ReportAgent for synthesis
β†’ Do NOT continue searching or gathering more evidence
β†’ The Judge has determined evidence quality is adequate
## Standard Workflow
1. SearchAgent: Find evidence from PubMed, ClinicalTrials.gov, and Europe PMC
2. HypothesisAgent: Generate mechanistic hypotheses (Drug -> Target -> Pathway -> Effect)
3. JudgeAgent: Evaluate if evidence is sufficient
4. If insufficient -> SearchAgent refines search based on gaps
5. If sufficient -> ReportAgent synthesizes final report
Focus on:
- Identifying specific molecular targets
- Understanding mechanism of action
- Finding clinical evidence supporting hypotheses
The final output should be a structured research report."""
def _get_progress_message(self, iteration: int) -> str:
"""Generate progress message with time estimation."""
rounds_remaining = max(self._max_rounds - iteration, 0)
est_seconds = rounds_remaining * self._EST_SECONDS_PER_ROUND
if est_seconds >= 60:
est_display = f"{est_seconds // 60}m {est_seconds % 60}s"
else:
est_display = f"{est_seconds}s"
return f"Round {iteration}/{self._max_rounds} (~{est_display} remaining)"
async def _init_workflow_events(self, query: str) -> AsyncGenerator[AgentEvent, None]:
"""Yield initialization events."""
yield AgentEvent(
type="started",
message=f"Starting research (Advanced mode): {query}",
iteration=0,
)
yield AgentEvent(
type="progress",
message="Loading embedding service (LlamaIndex/ChromaDB)...",
iteration=0,
)
async def _handle_timeout(self, iteration: int) -> AsyncGenerator[AgentEvent, None]:
"""Handle workflow timeout by attempting synthesis."""
logger.warning("Workflow timed out", iterations=iteration)
# ACTUALLY synthesize from gathered evidence
try:
from src.agents.magentic_agents import create_report_agent
from src.agents.state import get_magentic_state
state = get_magentic_state()
memory = state.memory
# Get evidence summary from memory
evidence_summary = await memory.get_context_summary()
# Create and invoke ReportAgent for synthesis
report_agent = create_report_agent(self._chat_client, domain=self.domain)
yield AgentEvent(
type="synthesizing",
message="Workflow timed out. Synthesizing available evidence...",
iteration=iteration,
)
# Invoke ReportAgent directly
# Note: ChatAgent.run() returns AgentRunResponse; access text via .text
synthesis_result = await report_agent.run(
"Synthesize research report from this evidence. "
f"If evidence is sparse, say so.\n\n{evidence_summary}"
)
yield AgentEvent(
type="complete",
message=synthesis_result.text,
data={"reason": "timeout_synthesis", "iterations": iteration},
iteration=iteration,
)
except Exception as synth_error:
logger.error("Timeout synthesis failed", error=str(synth_error))
yield AgentEvent(
type="complete",
message=(
f"Research timed out after {iteration} rounds. "
f"Evidence gathered but synthesis failed: {synth_error}"
),
data={"reason": "timeout_synthesis_failed", "iterations": iteration},
iteration=iteration,
)
async def run(self, query: str) -> AsyncGenerator[AgentEvent, None]:
"""
Run the workflow.
Args:
query: User's research question
Yields:
AgentEvent objects for real-time UI updates
"""
logger.info("Starting Advanced orchestrator", query=query)
async for event in self._init_workflow_events(query):
yield event
# Initialize context state
embedding_service = self._init_embedding_service()
yield AgentEvent(
type="progress",
message="Initializing research memory...",
iteration=0,
)
init_magentic_state(query, embedding_service)
yield AgentEvent(
type="progress",
message="Building agent team (Search, Judge, Hypothesis, Report)...",
iteration=0,
)
workflow = self._build_workflow()
task = self._create_task_prompt(query)
# UX FIX: Yield thinking state before blocking workflow call
# The workflow.run_stream() blocks for 2+ minutes on first LLM call
yield AgentEvent(
type="thinking",
message=(
f"Multi-agent reasoning in progress ({self._max_rounds} rounds max)... "
f"Estimated time: {self._max_rounds * 45 // 60}-"
f"{self._max_rounds * 60 // 60} minutes."
),
iteration=0,
)
iteration = 0
final_event_received = False
# ACCUMULATOR PATTERN: Track streaming content to bypass upstream Repr Bug
# Upstream bug in _magentic.py flattens message.contents and sets message.text
# to repr(message) if text is empty. We must reconstruct text from Deltas.
current_message_buffer: str = ""
current_agent_id: str | None = None
try:
async with asyncio.timeout(self._timeout_seconds):
async for event in workflow.run_stream(task):
# 1. Handle Streaming (Source of Truth for Content)
if isinstance(event, MagenticAgentDeltaEvent):
# Detect agent switch to clear buffer
if event.agent_id != current_agent_id:
current_message_buffer = ""
current_agent_id = event.agent_id
if event.text:
current_message_buffer += event.text
yield AgentEvent(
type="streaming",
message=event.text,
data={"agent_id": event.agent_id},
iteration=iteration,
)
continue
# 2. Handle Completion Signal
# We use our accumulated buffer instead of the corrupted event.message
if isinstance(event, MagenticAgentMessageEvent):
iteration += 1
comp_event, prog_event = self._handle_completion_event(
event, current_message_buffer, iteration
)
yield comp_event
yield prog_event
# Clear buffer after consuming
current_message_buffer = ""
continue
# 3. Handle other events normally
agent_event = self._process_event(event, iteration)
if agent_event:
if agent_event.type == "complete":
final_event_received = True
yield agent_event
# GUARANTEE: Always emit termination event if stream ends without one
# (e.g., max rounds reached)
if not final_event_received:
logger.warning(
"Workflow ended without final event",
iterations=iteration,
)
yield AgentEvent(
type="complete",
message=(
f"Research completed after {iteration} agent rounds. "
"Max iterations reached - results may be partial. "
"Try a more specific query for better results."
),
data={"iterations": iteration, "reason": "max_rounds_reached"},
iteration=iteration,
)
except TimeoutError:
async for event in self._handle_timeout(iteration):
yield event
except Exception as e:
logger.error("Workflow failed", error=str(e))
yield AgentEvent(
type="error",
message=f"Workflow error: {e!s}",
iteration=iteration,
)
def _handle_completion_event(
self, event: MagenticAgentMessageEvent, buffer: str, iteration: int
) -> tuple[AgentEvent, AgentEvent]:
"""Handle an agent completion event using the accumulated buffer."""
# Use buffer if available, otherwise fall back cautiously
# (Only fall back if buffer empty, which implies tool-only turn)
text_content = buffer
if not text_content:
# Try extraction but ignore repr strings AND empty strings
raw_text = self._extract_text(event.message)
if raw_text and not (raw_text.startswith("<") and "object at" in raw_text):
text_content = raw_text
else:
text_content = "Action completed (Tool Call)"
agent_name = event.agent_id or "unknown"
event_type = self._get_event_type_for_agent(agent_name)
completion_event = AgentEvent(
type=event_type,
message=f"{agent_name}: {text_content[:200]}...",
iteration=iteration,
)
# Progress update
rounds_remaining = max(self._max_rounds - iteration, 0)
est_seconds = rounds_remaining * 45
est_display = (
f"{est_seconds // 60}m {est_seconds % 60}s" if est_seconds >= 60 else f"{est_seconds}s"
)
progress_event = AgentEvent(
type="progress",
message=f"Round {iteration}/{self._max_rounds} (~{est_display} remaining)",
iteration=iteration,
)
return completion_event, progress_event
def _extract_text(self, message: Any) -> str:
"""
Defensively extract text from a message object.
Handles ChatMessage objects from both OpenAI and HuggingFace clients.
ChatMessage has: .text (str), .contents (list of content objects)
Also handles plain string messages (e.g., WorkflowOutputEvent.data).
"""
if not message:
return ""
# Priority 0: Handle plain string messages (e.g., WorkflowOutputEvent.data)
if isinstance(message, str):
# Filter out obvious repr-style noise
if not (message.startswith("<") and "object at" in message):
return message
return ""
# Priority 1: .text (standard ChatMessage text content)
if hasattr(message, "text") and message.text:
text = message.text
# Verify it's actually a string, not the object itself
if isinstance(text, str) and not (text.startswith("<") and "object at" in text):
return text
# Priority 2: .contents (list of FunctionCallContent, TextContent, etc.)
# This handles tool call responses from HuggingFace
if hasattr(message, "contents") and message.contents:
parts = []
for content in message.contents:
# TextContent has .text
if hasattr(content, "text") and content.text:
parts.append(str(content.text))
# FunctionCallContent has .name and .arguments
elif hasattr(content, "name"):
parts.append(f"[Tool: {content.name}]")
if parts:
return " ".join(parts)
# Priority 3: .content (legacy - some frameworks use singular)
if hasattr(message, "content") and message.content:
content = message.content
if isinstance(content, str):
return content
if isinstance(content, list):
return " ".join([str(c.text) for c in content if hasattr(c, "text")])
# Fallback: Return empty string instead of repr
# The repr is useless for display purposes
return ""
def _get_event_type_for_agent(
self, agent_name: str
) -> Literal["search_complete", "judge_complete", "hypothesizing", "synthesizing", "judging"]:
"""Map agent name to appropriate event type.
Args:
agent_name: The agent ID from the workflow event
Returns:
Event type string matching AgentEvent.type Literal
"""
agent_lower = agent_name.lower()
if "search" in agent_lower:
return "search_complete"
if "judge" in agent_lower:
return "judge_complete"
if "hypothes" in agent_lower:
return "hypothesizing"
if "report" in agent_lower:
return "synthesizing"
return "judging" # Default for unknown agents
def _smart_truncate(self, text: str, max_len: int = 200) -> str:
"""Truncate at sentence boundary to avoid cutting words."""
if len(text) <= max_len:
return text
# Find last sentence boundary before limit
truncated = text[:max_len]
last_period = truncated.rfind(". ")
if last_period > max_len // 2:
return truncated[: last_period + 1]
# Fallback to word boundary
return truncated.rsplit(" ", 1)[0] + "..."
def _process_event(self, event: Any, iteration: int) -> AgentEvent | None:
"""Process workflow event into AgentEvent."""
if isinstance(event, MagenticOrchestratorMessageEvent):
# FILTERING: Skip internal framework bookkeeping
if event.kind in ("task_ledger", "instruction"):
return None
# TRANSFORMATION: Handle user_task BEFORE text extraction
# (user_task uses static message, doesn't need text content)
if event.kind == "user_task":
return AgentEvent(
type="progress",
message="Manager assigning research task to agents...",
iteration=iteration,
)
# For other manager events, extract and validate text
text = self._extract_text(event.message)
if not text:
return None
# Default fallback for other manager events
return AgentEvent(
type="judging",
message=f"Manager ({event.kind}): {self._smart_truncate(text)}",
iteration=iteration,
)
# NOTE: MagenticAgentMessageEvent is handled in run() loop with Accumulator Pattern
# (see lines 322-335) and never reaches this method due to `continue` statement.
elif isinstance(event, MagenticFinalResultEvent):
text = self._extract_text(event.message) if event.message else "No result"
return AgentEvent(
type="complete",
message=text,
data={"iterations": iteration},
iteration=iteration,
)
# NOTE: MagenticAgentDeltaEvent is handled in run() loop with Accumulator Pattern
# (see lines 306-320) and never reaches this method due to `continue` statement.
elif isinstance(event, WorkflowOutputEvent):
if event.data:
# Use _extract_text to properly handle ChatMessage objects
text = self._extract_text(event.data)
return AgentEvent(
type="complete",
message=text if text else "Research complete (no synthesis)",
iteration=iteration,
)
return None
def _create_deprecated_alias() -> type["AdvancedOrchestrator"]:
"""Create a deprecated alias that warns on use."""
import warnings
class MagenticOrchestrator(AdvancedOrchestrator):
"""Deprecated alias for AdvancedOrchestrator.
.. deprecated:: 0.1.0
Use :class:`AdvancedOrchestrator` instead.
"""
def __init__(self, *args: Any, **kwargs: Any) -> None:
"""Initialize deprecated MagenticOrchestrator (use AdvancedOrchestrator)."""
warnings.warn(
"MagenticOrchestrator is deprecated, use AdvancedOrchestrator instead. "
"The name 'magentic' was confusing with the 'magentic' PyPI package.",
DeprecationWarning,
stacklevel=2,
)
super().__init__(*args, **kwargs)
return MagenticOrchestrator
# Backwards compatibility alias with deprecation warning
MagenticOrchestrator = _create_deprecated_alias()