File size: 28,593 Bytes
6b62834 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 | """ReAct Engine — the core Think → Act → Observe loop."""
import asyncio
import json
import time
import uuid
from typing import AsyncIterator, Optional
from agentic_rag.data.models import (
AgentEvent,
AgentEventType,
AgentInput,
AgentOutput,
LLMChunk,
LLMResponse,
Message,
ToolCall,
ToolCallResult,
ToolDefinition,
)
from agentic_rag.agent.react_parser import (
ReActStep,
extract_final_answer,
format_observation,
parse_react_output,
)
from agentic_rag.agent.react_prompt import build_react_prompt, build_tools_description
from agentic_rag.services.llm.base import (
BaseLLMProvider,
ReasoningStreamFilter,
strip_reasoning,
)
class ReActEngine:
"""ReAct (Reasoning + Acting) reasoning engine.
Executes the Think → Act → Observe loop that powers the agent.
"""
def __init__(
self,
llm: BaseLLMProvider,
tools: list,
system_prompt_template: str,
max_iterations: int = 10,
stop_on_error: bool = False,
enable_native_tool_calls: bool = True,
require_tool_call: bool = False,
):
"""
Args:
llm: LLM provider for generation.
tools: List of BaseTool instances available to the agent.
system_prompt_template: Template string for the system prompt.
max_iterations: Maximum ReAct loop iterations.
stop_on_error: If True, stop on first tool error.
enable_native_tool_calls: If False, tools are NOT sent to the LLM
(pure ReAct text mode for models without function calling).
require_tool_call: If True, reject a final answer until at least one
available tool has completed successfully. Used for freshness-
sensitive queries where model memory is not acceptable evidence.
"""
self.llm = llm
self.tools = tools
self.system_prompt_template = system_prompt_template
self.max_iterations = max_iterations
self.stop_on_error = stop_on_error
self.enable_native_tool_calls = enable_native_tool_calls
self.require_tool_call = require_tool_call
self._tool_map = {t.name: t for t in tools}
async def run(self, input: AgentInput, turn_id: str = "") -> AgentOutput:
"""Execute the ReAct loop (non-streaming)."""
if not turn_id:
turn_id = uuid.uuid4().hex
messages = self._build_initial_messages(input)
tool_calls_made: list[ToolCallResult] = []
total_usage: dict[str, int] = {"prompt_tokens": 0, "completion_tokens": 0}
invalid_output_count = 0 # Track consecutive invalid outputs
max_invalid_attempts = 2 # Force final answer after 2 consecutive invalid outputs
executed_sigs: set[str] = set() # (tool, args) dedupe — mirrors stream()
for iteration in range(self.max_iterations):
# Contract for the NEXT turn: a native function-call round ends with
# "continue", so if no new tool result was appended after this point
# the model must answer in text ReAct format (never emit format-only
# text alongside a function call, or the final tool result gets
# ignored and the last observation is lost).
n_msgs_before = len(messages)
messages.append(Message.user(
"【重要】如果你决定调用工具,请只发起工具调用,不要同时输出任何 Thought/Action 格式行;"
"如果你不调用工具,请按格式输出:Thought: ... Final Answer: ...(或 Thought/Action/Action Input 发起文本式工具调用)。"
))
response = await self.llm.agenerate(messages, self._get_llm_tool_definitions())
if len(messages) > n_msgs_before:
messages.pop() # remove the contract — don't pollute history
total_usage["prompt_tokens"] += response.usage.get("prompt_tokens", 0)
total_usage["completion_tokens"] += response.usage.get("completion_tokens", 0)
# Handle native tool calls (from providers that support function calling)
if response.tool_calls:
invalid_output_count = 0 # Reset on valid tool call
for tc in response.tool_calls:
args = tc.arguments or {}
if not args:
messages.append(Message.tool(
content=f"Error: '{tc.name}' 缺少参数",
tool_call_id=tc.id,
))
continue
sig = (tc.name, json.dumps(args, sort_keys=True, ensure_ascii=False))
if sig in executed_sigs:
messages.append(Message.tool(
content=f"Error: 禁止重复调用 '{tc.name}'(参数相同)。请基于已有 Observation 直接输出 Final Answer。",
tool_call_id=tc.id,
))
continue
executed_sigs.add(sig)
result = await self._execute_tool(tc.name, args)
tool_calls_made.append(result)
messages.append(Message.tool(
content=str(result.result) if not result.error else f"Error: {result.error}",
tool_call_id=tc.id,
))
continue
# Parse ReAct format output
step = parse_react_output(response.content, list(self._tool_map.keys()))
if step.is_final:
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
invalid_output_count += 1
messages.append(Message.assistant(strip_reasoning(response.content)))
messages.append(Message.user(
"该问题必须先调用可用的网络搜索工具并获得成功的 Observation。"
"不得依据模型记忆直接回答,也不得自行生成引用。请立即调用工具。"
))
continue
return AgentOutput(
messages=messages,
final_answer=step.final_answer,
tool_calls_made=tool_calls_made,
usage=total_usage,
iterations=iteration + 1,
)
if step.action:
invalid_output_count = 0 # Reset on valid action
# Guard: skip if action_input is empty (broken parse)
if not step.action_input:
messages.append(Message.user(
f"'{step.action}' 需要参数,请在 Action Input 中提供 JSON。"
))
continue
sig = (step.action, json.dumps(step.action_input, sort_keys=True, ensure_ascii=False))
if sig in executed_sigs:
messages.append(Message.user(
f"你已经用相同参数调用过 '{step.action}' 了。请基于已有 Observation 直接输出 Final Answer。"
))
continue
executed_sigs.add(sig)
# Execute the tool
result = await self._execute_tool(step.action, step.action_input)
tool_calls_made.append(result)
# Format observation and append to messages
observation = format_observation(
step.action,
str(result.result) if result.result else "",
result.error,
)
# Append the assistant's ReAct output + observation as a single message
messages.append(Message.assistant(response.content))
messages.append(Message.user(observation))
else:
# No valid action and no final answer — LLM output is unparseable
invalid_output_count += 1
if invalid_output_count >= max_invalid_attempts:
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
return AgentOutput(
messages=messages,
final_answer=self._live_search_failure_answer(),
tool_calls_made=tool_calls_made,
usage=total_usage,
iterations=iteration + 1,
)
# Force exit after repeated invalid outputs to prevent infinite loop
final = extract_final_answer(response.content)
if not final:
# Use accumulated content as fallback answer
final = response.content.strip() or "抱歉,我暂时无法回答这个问题。"
return AgentOutput(
messages=messages,
final_answer=final,
tool_calls_made=tool_calls_made,
usage=total_usage,
iterations=iteration + 1,
)
# Prompt LLM to continue with correct format
messages.append(Message.user(
"你的输出格式不正确。请严格按照以下格式之一输出:\n"
"1. 调用工具:Thought: ...\nAction: tool_name\nAction Input: {\"param\": \"value\"}\n"
"2. 最终答案:Thought: ...\nFinal Answer: ..."
))
# Max iterations reached
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
final = self._live_search_failure_answer()
else:
final = await self._force_final_answer(messages)
return AgentOutput(
messages=messages,
final_answer=final,
tool_calls_made=tool_calls_made,
usage=total_usage,
iterations=self.max_iterations,
)
async def stream(self, input: AgentInput, turn_id: str = "") -> AsyncIterator[AgentEvent]:
"""Execute the ReAct loop with streaming events."""
if not turn_id:
turn_id = uuid.uuid4().hex
messages = self._build_initial_messages(input)
tool_calls_made: list[ToolCallResult] = []
invalid_output_count = 0 # Track consecutive invalid outputs
max_invalid_attempts = 2 # Force final answer after 2 consecutive invalid outputs
executed_sigs: set[tuple[str, str]] = set()
for iteration in range(self.max_iterations):
# Emit thought event
yield AgentEvent(
event_type=AgentEventType.THOUGHT,
data={"iteration": iteration},
turn_id=turn_id,
)
# Stream LLM generation — collect both text deltas AND native tool calls.
full_content = ""
# Native function-calling state (accumulated across streaming chunks)
native_tool_name = ""
native_tool_args = ""
has_native_tool_call = False
stream_error = None
# Incremental reasoning stripper — thinking models (<think> blocks)
# must not leak their reasoning to the UI or into the ReAct parser.
think_filter = ReasoningStreamFilter()
# Answer gating: only stream the actual answer text to the UI.
# "Thought:" lines and any pre-answer monologue are buffered and
# dropped — the final answer is extracted from full_content below.
pending_delta = ""
answer_streaming = False
try:
async for chunk in self.llm.agenerate_stream(messages, self._get_llm_tool_definitions()):
if chunk.content_delta:
delta = think_filter.feed(chunk.content_delta)
full_content += delta
# After tool call detection, further text is likely
# post-tool narration — suppress to reduce noise.
if not delta or has_native_tool_call:
continue
if answer_streaming:
emit = delta
else:
pending_delta += delta
if "Final Answer:" in pending_delta:
emit = pending_delta.split("Final Answer:", 1)[1].lstrip()
answer_streaming = True
else:
continue
if emit:
yield AgentEvent(
event_type=AgentEventType.TEXT_DELTA,
data={"content": emit},
turn_id=turn_id,
)
# Collect native function-call deltas (Qwen / OpenAI function calling)
if chunk.tool_call_delta:
has_native_tool_call = True
if chunk.tool_call_delta.get("name"):
native_tool_name = chunk.tool_call_delta["name"]
if chunk.tool_call_delta.get("arguments"):
native_tool_args += chunk.tool_call_delta["arguments"]
# Flush text held back for partial-tag detection
tail = think_filter.flush()
if tail:
full_content += tail
if not has_native_tool_call:
if answer_streaming:
yield AgentEvent(
event_type=AgentEventType.TEXT_DELTA,
data={"content": tail},
turn_id=turn_id,
)
else:
pending_delta += tail
if "Final Answer:" in pending_delta:
emit = pending_delta.split("Final Answer:", 1)[1].lstrip()
if emit:
answer_streaming = True
yield AgentEvent(
event_type=AgentEventType.TEXT_DELTA,
data={"content": emit},
turn_id=turn_id,
)
except Exception as e:
stream_error = str(e)
import sys
print(f" [ReAct] ⚠ LLM stream error (iteration {iteration}): {e}", flush=True)
sys.stdout.flush()
# ── Handle stream error ──
if stream_error and not full_content.strip():
yield AgentEvent(
event_type=AgentEventType.ERROR,
data={"error": f"LLM stream failed: {stream_error}"},
turn_id=turn_id,
)
yield AgentEvent(
event_type=AgentEventType.DONE,
data={"final_answer": f"抱歉,模型服务连接中断:{stream_error},请稍后重试。"},
turn_id=turn_id,
)
return
# ── Resolve action: native function calling takes priority ──
if has_native_tool_call and native_tool_name:
invalid_output_count = 0 # Reset on valid tool call
try:
action_input = json.loads(native_tool_args) if native_tool_args else {}
except json.JSONDecodeError:
action_input = {"query": native_tool_args} if native_tool_args else {}
sig = (native_tool_name, json.dumps(action_input, sort_keys=True, ensure_ascii=False))
if sig in executed_sigs:
messages.append(Message.user(
f"你已经用相同参数调用过 '{native_tool_name}'。请根据已有 Observation 输出 Final Answer。"
))
continue
executed_sigs.add(sig)
yield AgentEvent(
event_type=AgentEventType.TOOL_CALL_START,
data={"tool": native_tool_name, "input": action_input},
turn_id=turn_id,
)
result = await self._execute_tool(native_tool_name, action_input)
tool_calls_made.append(result)
yield AgentEvent(
event_type=AgentEventType.TOOL_CALL_RESULT,
data={
"tool": native_tool_name,
"success": not result.error,
"result": str(result.result)[:500] if result.result else "",
"error": result.error,
},
turn_id=turn_id,
)
observation = format_observation(
native_tool_name,
str(result.result) if result.result else "",
result.error,
)
messages.append(Message.assistant(
strip_reasoning(full_content).strip()
or f"Thought: 调用 {native_tool_name}\nAction: {native_tool_name}\nAction Input: {json.dumps(action_input, ensure_ascii=False)}"
))
messages.append(Message.user(observation))
continue
# ── Fallback: parse ReAct text format ──
step = parse_react_output(full_content, list(self._tool_map.keys()))
if step.is_final:
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
invalid_output_count += 1
messages.append(Message.assistant(strip_reasoning(full_content)))
messages.append(Message.user(
"该问题必须先调用可用的网络搜索工具并获得成功的 Observation。"
"不得依据模型记忆直接回答,也不得自行生成引用。请立即调用工具。"
))
continue
yield AgentEvent(
event_type=AgentEventType.DONE,
data={"final_answer": step.final_answer, "iterations": iteration + 1},
turn_id=turn_id,
)
return
if step.action:
invalid_output_count = 0 # Reset on valid action
# Guard: if action_input is empty and the tool requires args, skip
if not step.action_input:
messages.append(Message.user(
f"'{step.action}' 需要参数,请在 Action Input 中提供 JSON。"
))
continue
sig = (step.action, json.dumps(step.action_input, sort_keys=True, ensure_ascii=False))
if sig in executed_sigs:
messages.append(Message.user(
f"你已经用相同参数调用过 '{step.action}'。请根据已有 Observation 输出 Final Answer。"
))
continue
executed_sigs.add(sig)
yield AgentEvent(
event_type=AgentEventType.TOOL_CALL_START,
data={"tool": step.action, "input": step.action_input},
turn_id=turn_id,
)
result = await self._execute_tool(step.action, step.action_input)
tool_calls_made.append(result)
yield AgentEvent(
event_type=AgentEventType.TOOL_CALL_RESULT,
data={
"tool": step.action,
"success": not result.error,
"result": str(result.result)[:500] if result.result else "",
"error": result.error,
},
turn_id=turn_id,
)
observation = format_observation(
step.action,
str(result.result) if result.result else "",
result.error,
)
messages.append(Message.assistant(strip_reasoning(full_content)))
messages.append(Message.user(observation))
else:
# No valid action and no final answer — LLM output is unparseable
invalid_output_count += 1
if invalid_output_count >= max_invalid_attempts:
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
yield AgentEvent(
event_type=AgentEventType.DONE,
data={"final_answer": self._live_search_failure_answer(),
"iterations": iteration + 1},
turn_id=turn_id,
)
return
# Force exit after repeated invalid outputs to prevent infinite loop
final = extract_final_answer(full_content)
if not final:
final = await self._force_final_answer(messages)
if not final or ("Thought:" in final and "Action" in final):
final = full_content.strip()
yield AgentEvent(
event_type=AgentEventType.DONE,
data={"final_answer": final or "抱歉,我暂时无法回答这个问题。",
"iterations": iteration + 1},
turn_id=turn_id,
)
return
messages.append(Message.user(
"你的输出格式不正确。请严格按照以下格式之一输出:\n"
"1. 调用工具:Thought: ...\nAction: tool_name\nAction Input: {\"param\": \"value\"}\n"
"2. 最终答案:Thought: ...\nFinal Answer: ..."
))
# Max iterations — force LLM to give a final answer with what it has
if self.require_tool_call and not self._has_successful_tool_result(tool_calls_made):
final = self._live_search_failure_answer()
else:
final = await self._force_final_answer(messages)
yield AgentEvent(
event_type=AgentEventType.DONE,
data={"final_answer": final or "Max iterations reached. Could not complete the task.",
"iterations": self.max_iterations},
turn_id=turn_id,
)
def _build_initial_messages(self, input: AgentInput) -> list[Message]:
"""Build the initial message list for the ReAct loop."""
tools_desc = build_tools_description(self._get_tool_definitions())
memory_context = self._format_messages(input.messages)
system_prompt = build_react_prompt(
tools_description=tools_desc,
memory_context=memory_context,
)
messages = [Message.system(system_prompt)]
# Check if input already has a multimodal user message
has_multimodal_query = any(
isinstance(m.content, list) and m.role.value == "user"
for m in input.messages
)
# Add conversation history (excluding system messages)
for msg in input.messages:
if msg.role.value != "system":
messages.append(msg)
# Add current query — skip if already included as multimodal message
if not has_multimodal_query:
query = input.query
if input.multimodal and input.multimodal.text:
query = input.multimodal.text
messages.append(Message.user(query))
return messages
async def _execute_tool(self, name: str, arguments: dict) -> ToolCallResult:
"""Execute a tool by name with arguments.
Falls back to the global tool registry when the tool isn't in this
engine's filtered set: the model sometimes emits a tool it knows from
context (e.g. an MCP tool) that the router didn't hand it, and
answering "tool not found" wastes a turn when the tool actually
exists in the process.
"""
call_id = f"call_{uuid.uuid4().hex[:12]}"
tool = self._tool_map.get(name)
if tool is None:
try:
from agentic_rag.orchestration.l1_tools.registry import get_tool_registry
tool = get_tool_registry().get(name)
except Exception:
tool = None
if tool is None:
return ToolCallResult(
call_id=call_id,
name=name,
result=None,
error=f"Tool '{name}' not found. Available: {list(self._tool_map.keys())}",
)
try:
# Add timeout to prevent hanging on slow tools
result = await asyncio.wait_for(tool.execute(**arguments), timeout=60.0)
return ToolCallResult(
call_id=call_id,
name=name,
result=result,
)
except asyncio.TimeoutError:
return ToolCallResult(
call_id=call_id,
name=name,
result=None,
error=f"Tool '{name}' execution timed out after 60 seconds",
)
except Exception as e:
return ToolCallResult(
call_id=call_id,
name=name,
result=None,
error=str(e),
)
@staticmethod
def _has_successful_tool_result(results: list[ToolCallResult]) -> bool:
"""Return whether at least one tool produced usable evidence."""
return any(not result.error and result.result for result in results)
@staticmethod
def _live_search_failure_answer() -> str:
"""Fail closed when fresh information could not be retrieved."""
return (
"当前问题需要实时网络信息,但本次未能获得有效的搜索结果,"
"因此无法可靠确认。请稍后重试;为避免误导,我不会使用模型记忆猜测答案或编造来源。"
)
def _get_tool_definitions(self) -> list[ToolDefinition]:
"""Get all tool definitions for prompt construction and parsing."""
return [t.to_definition() for t in self.tools]
def _get_llm_tool_definitions(self) -> list[ToolDefinition]:
"""Get definitions passed through the provider's native tools API."""
if not self.enable_native_tool_calls:
return []
return self._get_tool_definitions()
@staticmethod
def _format_messages(messages: list[Message]) -> str:
"""Format conversation history for the prompt."""
if not messages:
return ""
lines = []
for msg in messages[-10:]: # Last 10 messages
content = msg.content
if isinstance(content, list):
# Extract text parts for history summary
texts = [p.get("text", "") for p in content if isinstance(p, dict) and p.get("text")]
img_count = sum(1 for p in content if isinstance(p, dict) and p.get("type") == "image_url")
parts = texts
if img_count:
parts.append(f"[{img_count} image(s)]")
content = " ".join(parts) if parts else "[multimodal content]"
lines.append(f"{msg.role.value}: {str(content)[:200]}")
return "\n".join(lines)
async def _force_final_answer(self, messages: list[Message]) -> str:
"""Force the LLM to produce a final answer when max iterations are reached."""
messages.append(Message.user(
"已达到最大步数。请基于已有信息给出 Final Answer。"
))
response = await self.llm.agenerate(messages)
final = extract_final_answer(response.content)
return final or response.content
|