Spaces:
Running
Running
File size: 22,068 Bytes
3e26b02 1d7a836 3e26b02 81840a0 3e26b02 1d7a836 3e26b02 1d7a836 3e26b02 1d7a836 fd366ed 71f0b10 3e26b02 1d7a836 3e26b02 1d7a836 3e26b02 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 | import os
import yaml
import asyncio
import inspect
import json
import time
from typing import Any, Dict, AsyncIterator, List, Tuple, Optional
import gradio as gr
from dotenv import load_dotenv
from retrac.graph import build_graph
# -----------------------------
# Your existing logic (kept)
# -----------------------------
def load_config(config_path: str) -> dict:
with open(config_path, "r", encoding="utf-8") as f:
return yaml.safe_load(f)
async def stream_graph_execution(
graph,
initial_state: Dict[str, Any],
) -> AsyncIterator[Dict[str, Any]]:
compiled_graph = graph.compile()
recursion_limit = int(os.getenv("RECURSION_LIMIT", "10000"))
run_config = {"recursion_limit": recursion_limit}
state = initial_state.copy()
async for event in compiled_graph.astream(state, config=run_config):
yield event
# same as your code: merge node outputs into state
for node_name, node_output in event.items():
if node_output is not None:
state.update(node_output)
async def run_streaming(config_path: str, question: str) -> AsyncIterator[Dict[str, Any]]:
cfg = load_config(config_path)
graph = build_graph(cfg)
initial_state = {"question": question}
async for event in stream_graph_execution(graph, initial_state):
yield event
# -----------------------------
# Message serialization helpers
# -----------------------------
def _safe_str(x: Any) -> str:
try:
return str(x)
except Exception:
return repr(x)
def _strip_think_tags(text: str) -> str:
if not text:
return text
return (
text.replace("<think>", "")
.replace("</think>", "")
.replace("<think/>", "")
.strip()
)
def _try_parse_json(value: Any) -> Any:
if isinstance(value, (dict, list)):
return value
if isinstance(value, str):
try:
return json.loads(value)
except Exception:
return value
return value
def _format_json_block(value: Any) -> str:
parsed = _try_parse_json(value)
if isinstance(parsed, (dict, list)):
body = json.dumps(parsed, ensure_ascii=False, indent=2)
else:
body = _safe_str(parsed)
return f"```json\n{body}\n```"
def _debug_log(message: str, data: Dict[str, Any], hypothesis_id: str, run_id: str = "pre-fix") -> None:
# #region agent log
try:
payload = {
"id": f"log_{int(time.time() * 1000)}_{os.getpid()}",
"timestamp": int(time.time() * 1000),
"location": "RE-TRAC/app.py",
"message": message,
"data": data,
"sessionId": "debug-session",
"runId": run_id,
"hypothesisId": hypothesis_id,
}
with open("/home/azureuser/jialiang/repo/tz-dr/.cursor/debug.log", "a", encoding="utf-8") as f:
f.write(json.dumps(payload, ensure_ascii=True) + "\n")
except Exception:
pass
# #endregion
CHATBOT_EXPECTS_MESSAGES: Optional[bool] = None
CHATBOT_SUPPORTS_GROUPING: Optional[bool] = None
def _gradio_version_major() -> Optional[int]:
version = getattr(gr, "__version__", None)
if not version:
return None
try:
return int(str(version).split(".")[0])
except Exception:
return None
def _message_fingerprint(msg: Any) -> str:
role, content = _extract_role_and_content(msg)
return f"{role}::{content}"
def _is_tool_message(msg: Any) -> bool:
if isinstance(msg, dict):
return (msg.get("role") == "tool") or (msg.get("type") == "tool")
return getattr(msg, "type", None) == "tool"
def _extract_role_and_content(msg: Any) -> Tuple[str, str]:
"""
Convert various message types into (role, content) for display.
Supports:
- LangChain BaseMessage-like objects (has .type/.content)
- dict messages {"role": "...", "content": "..."} or similar
"""
# dict-like
if isinstance(msg, dict):
role = msg.get("role") or msg.get("type") or "unknown"
content = msg.get("content")
if content is None:
# fallback: show whole dict
content = _safe_str(msg)
return str(role), _safe_str(content)
# BaseMessage-like
mtype = getattr(msg, "type", None)
content = getattr(msg, "content", None)
# ToolMessage often has name/tool_call_id, show them if present
if mtype == "tool":
return "tool", _safe_str(content)
if mtype == "human":
return "user", _safe_str(content)
if mtype == "ai":
return "assistant", _safe_str(content)
if mtype == "system":
return "system", _safe_str(content)
# unknown object: try best
if content is not None and mtype is not None:
return _safe_str(mtype), _safe_str(content)
return "unknown", _safe_str(msg)
def _extract_tool_calls(msg: Any) -> List[Dict[str, Any]]:
if isinstance(msg, dict):
tool_calls = msg.get("tool_calls") or msg.get("tool_calls", [])
if isinstance(tool_calls, list):
return tool_calls
return []
tool_calls = getattr(msg, "tool_calls", None)
if isinstance(tool_calls, list):
return tool_calls
return []
def _split_ai_think(content: str) -> List[str]:
if not content:
return []
if "</think>" not in content and "<think/>" not in content and "<think>" not in content:
return [content]
if "<think/>" in content and "</think>" not in content:
after = _strip_think_tags(content.replace("<think/>", "", 1))
return [after] if after else []
if "<think>" in content and "</think>" not in content:
cleaned = _strip_think_tags(content)
return [cleaned] if cleaned else []
before, after = content.split("</think>", 1)
before = _strip_think_tags(before)
after = _strip_think_tags(after.lstrip())
parts = []
if before:
parts.append(before)
if after:
parts.append(after)
return parts
def _tool_call_to_text(tool_call: Any) -> str:
if isinstance(tool_call, dict):
name = tool_call.get("name") or tool_call.get("tool") or "tool_call"
args = tool_call.get("args") or tool_call.get("arguments") or {}
else:
name = getattr(tool_call, "name", None) or getattr(tool_call, "tool", None) or "tool_call"
args = getattr(tool_call, "args", None) or getattr(tool_call, "arguments", None) or {}
return f"**Tool Call** ({name})\n{_format_json_block(args)}"
def _explode_message(msg: Any, is_nonfirst_human: bool) -> List[Tuple[str, str, str]]:
role, content = _extract_role_and_content(msg)
parts: List[Tuple[str, str, str]] = []
render_role = "assistant" if role == "user" else role
_debug_log(
"explode_message entry",
{
"role": role,
"render_role": render_role,
"has_think": ("<think" in content) or ("</think>" in content),
"content_preview": _safe_str(content)[:200],
},
"H8",
)
if _is_tool_message(msg):
tool_body = _format_json_block(content)
return [(render_role, f"**Tool**\n{tool_body}", "tool")]
if role in ("assistant", "user"):
split_parts = _split_ai_think(content)
_debug_log(
"explode_message split",
{
"role": role,
"split_count": len(split_parts),
"split_previews": [p[:120] for p in split_parts],
},
"H8",
)
has_think = ("</think>" in content) or ("<think/>" in content) or ("<think>" in content)
for idx, part in enumerate(split_parts):
if has_think:
label = "**Model Reasoning**" if idx == 0 and "</think>" in content else "**Model Output**"
body = f"{label}\n{_strip_think_tags(part)}"
else:
body = _strip_think_tags(part)
if role == "user" and is_nonfirst_human:
body = f"{body}"
parts.append((render_role, body, role))
for tool_call in _extract_tool_calls(msg):
parts.append((render_role, _tool_call_to_text(tool_call), "tool_call"))
return parts or [(render_role, content, role)]
return [(role, content, "assistant" if role in ("assistant", "system") else "user")]
def _explode_messages(messages: List[Any]) -> List[Tuple[str, str, str]]:
exploded: List[Tuple[str, str, str]] = []
human_count = 0
for msg in messages:
role, _ = _extract_role_and_content(msg)
if role == "user":
human_count += 1
is_nonfirst_human = human_count > 1
else:
is_nonfirst_human = False
exploded.extend(_explode_message(msg, is_nonfirst_human))
return exploded
def serialize_messages_for_gradio(messages: List[Any]) -> List[Dict[str, str]]:
"""
Gradio Chatbot(type="messages") expects:
[{"role": "user"/"assistant"/"system", "content": "..."}, ...]
Tool messages are mapped into role="assistant" with a prefix.
"""
out: List[Dict[str, str]] = []
for role, content, kind in _explode_messages(messages):
is_tool = kind in ("tool_call", "tool")
# normalize role into gradio-friendly roles
if role not in ("user", "assistant", "system"):
# put everything else as assistant, but keep a prefix
content = f"{content}"
role = "assistant"
out.append({"role": role, "content": content})
return out
def _chatbot_supports_messages() -> bool:
try:
params = list(inspect.signature(gr.Chatbot).parameters.keys())
supports = "type" in params
grouping = "group_consecutive_messages" in params
major = _gradio_version_major()
_debug_log(
"chatbot signature check",
{
"supports_messages": supports,
"supports_grouping": grouping,
"params": params,
"gradio_version": getattr(gr, "__version__", None),
"gradio_major": major,
},
"H1",
)
global CHATBOT_SUPPORTS_GROUPING
CHATBOT_SUPPORTS_GROUPING = grouping
return supports
except Exception:
_debug_log(
"chatbot signature check failed",
{"supports_messages": False},
"H1",
)
return False
def _should_use_messages_format() -> bool:
major = _gradio_version_major()
if major is not None and major >= 6:
_debug_log(
"format decision",
{"reason": "gradio_major>=6", "gradio_major": major},
"H1",
)
return True
if CHATBOT_SUPPORTS_MESSAGES:
_debug_log(
"format decision",
{"reason": "chatbot_supports_type", "gradio_major": major},
"H1",
)
return True
_debug_log(
"format decision",
{"reason": "legacy_tuple", "gradio_major": major},
"H1",
)
return False
CHATBOT_SUPPORTS_MESSAGES = _chatbot_supports_messages()
def serialize_messages_for_legacy_chatbot(
messages: List[Any],
) -> List[Tuple[Optional[str], Optional[str]]]:
"""
Legacy gr.Chatbot expects list of (user, bot) tuples.
We map non-user roles onto the bot side with a role prefix.
"""
rendered: List[Tuple[Optional[str], Optional[str]]] = []
for role, content, kind in _explode_messages(messages):
is_tool = kind in ("tool_call", "tool")
if role == "user":
rendered.append((None, f"**RE-TRAC: REcursive TRAjectory Compression**\n{content}"))
continue
if role != "assistant":
content = f"{content}"
if is_tool:
rendered.append((None, content))
else:
if rendered and rendered[-1][1] is None:
rendered[-1] = (rendered[-1][0], content)
else:
rendered.append((None, content))
return rendered
def serialize_messages_for_chatbot(messages: List[Any]) -> List[Any]:
_debug_log(
"serialize_messages_for_chatbot",
{
"supports_messages": CHATBOT_SUPPORTS_MESSAGES,
"expects_messages": CHATBOT_EXPECTS_MESSAGES,
"count": len(messages),
},
"H2",
)
if CHATBOT_EXPECTS_MESSAGES:
return serialize_messages_for_gradio(messages)
return serialize_messages_for_legacy_chatbot(messages)
def append_messages_for_chatbot(
rendered: List[Any],
tail_messages: List[Any],
) -> List[Any]:
_debug_log(
"append_messages_for_chatbot",
{
"supports_messages": CHATBOT_SUPPORTS_MESSAGES,
"expects_messages": CHATBOT_EXPECTS_MESSAGES,
"rendered_len": len(rendered),
"tail_len": len(tail_messages),
},
"H3",
)
if CHATBOT_EXPECTS_MESSAGES:
rendered.extend(serialize_messages_for_gradio(tail_messages))
return rendered
tail_rendered = serialize_messages_for_legacy_chatbot(tail_messages)
if not tail_rendered:
return rendered
if not rendered:
return tail_rendered
first = tail_rendered[0]
if rendered[-1][1] is None and first[0] is None and first[1] is not None:
rendered[-1] = (rendered[-1][0], first[1])
tail_rendered = tail_rendered[1:]
rendered.extend(tail_rendered)
return rendered
def fingerprint_messages(messages: List[Any]) -> List[str]:
"""
Create a stable-ish fingerprint list for prefix checks.
We avoid id() because objects may be recreated.
"""
fp: List[str] = []
for m in messages:
role, content = _extract_role_and_content(m)
fp.append(f"{role}::{content}")
return fp
def compute_render_strategy(
prev_fp: List[str],
new_fp: List[str],
) -> Tuple[str, int]:
"""
Decide whether we can append incrementally or must rebuild.
Returns:
("append", start_index) if prev is a prefix of new
("rebuild", 0) otherwise
"""
if not prev_fp:
return ("rebuild", 0)
if len(new_fp) >= len(prev_fp) and new_fp[: len(prev_fp)] == prev_fp:
return ("append", len(prev_fp))
return ("rebuild", 0)
# -----------------------------
# Gradio app
# -----------------------------
CONFIG_PATH = os.getenv("RETRAC_CONFIG", "retrac/30B.yaml")
async def run_once(query: str, ui_state: Optional[Dict[str, Any]]):
"""
Single-run streaming handler.
ui_state stores:
- "prev_fp": fingerprint list for diffing
- "rendered": messages already rendered (as gradio 'messages' format)
- "merged_state": current merged graph state (dict)
- "done": bool
"""
if ui_state is None:
ui_state = {}
_debug_log(
"run_once entry",
{
"query": _safe_str(query),
"done": bool(ui_state.get("done")),
"supports_messages": CHATBOT_SUPPORTS_MESSAGES,
"expects_messages": CHATBOT_EXPECTS_MESSAGES,
},
"H4",
)
# Enforce one-time input per run.
if ui_state.get("done"):
# already finished; no second run without reset
yield (
ui_state.get("rendered", []),
ui_state,
gr.update(interactive=False),
gr.update(interactive=False),
"Already finished. Click Reset to run again.",
)
return
merged_state: Dict[str, Any] = ui_state.get("merged_state", {})
prev_fp: List[str] = ui_state.get("prev_fp", [])
seen_fp: List[str] = ui_state.get("seen_fp", [])
all_messages: List[Any] = ui_state.get("all_messages", [])
rendered: List[Dict[str, str]] = ui_state.get("rendered", [])
# disable input immediately after submit
yield (
rendered,
ui_state,
gr.update(interactive=False),
gr.update(interactive=False),
"Running...",
)
try:
async for event in run_streaming(CONFIG_PATH, query):
# merge node outputs into merged_state (same logic as your runner)
for _, node_output in event.items():
if node_output is not None:
merged_state.update(node_output)
# pick the full message list from merged_state
full_messages = merged_state.get("messages", [])
# accumulate messages; never clear already displayed ones
new_messages: List[Any] = []
for m in full_messages:
fp = _message_fingerprint(m)
if fp not in seen_fp:
seen_fp.append(fp)
all_messages.append(m)
new_messages.append(m)
if not rendered:
rendered = serialize_messages_for_chatbot(all_messages)
elif new_messages:
rendered = append_messages_for_chatbot(rendered, new_messages)
new_fp = fingerprint_messages(full_messages)
_debug_log(
"rendered update",
{
"mode": "accumulate",
"rendered_len": len(rendered),
"rendered_item_type": type(rendered[0]).__name__ if rendered else None,
"all_messages_len": len(all_messages),
"new_messages_len": len(new_messages),
},
"H5",
)
prev_fp = new_fp
# update ui_state
ui_state["merged_state"] = merged_state
ui_state["prev_fp"] = prev_fp
ui_state["seen_fp"] = seen_fp
ui_state["all_messages"] = all_messages
ui_state["rendered"] = rendered
# show some lightweight status
status = "Running..."
if "final" in event:
status = "Finished."
ui_state["done"] = True
yield (
rendered,
ui_state,
gr.update(interactive=False),
gr.update(interactive=False),
status,
)
if "final" in event:
break
except Exception as e:
ui_state["done"] = True
yield (
rendered,
ui_state,
gr.update(interactive=False),
gr.update(interactive=False),
f"Error: {_safe_str(e)}",
)
def reset():
# re-enable input
return [], None, gr.update(interactive=True, value=""), gr.update(interactive=True), "Idle."
with gr.Blocks() as demo:
gr.Markdown("## RE-TRAC (REcursive TRAjectory Compression) 30B Demo")
with gr.Row():
with gr.Column(scale=6):
gr.Markdown(f"Config: `{CONFIG_PATH}`")
chatbot_kwargs = {"height": 520, "label": "Messages (full stream)"}
if CHATBOT_SUPPORTS_GROUPING:
chatbot_kwargs["group_consecutive_messages"] = False
if CHATBOT_SUPPORTS_MESSAGES:
chatbot = gr.Chatbot(type="messages", **chatbot_kwargs)
else:
chatbot = gr.Chatbot(**chatbot_kwargs)
# #region agent log
try:
attr_type = getattr(chatbot, "type", None)
attr_format = getattr(chatbot, "format", None)
attr_data_format = getattr(chatbot, "data_format", None)
CHATBOT_EXPECTS_MESSAGES = _should_use_messages_format() or (
(attr_type == "messages") or (attr_format == "messages") or (attr_data_format == "messages")
)
_debug_log(
"chatbot instance attrs",
{
"attr_type": attr_type,
"attr_format": attr_format,
"attr_data_format": attr_data_format,
"expects_messages": CHATBOT_EXPECTS_MESSAGES,
},
"H1",
)
except Exception:
pass
# #endregion
status = gr.Markdown("Idle.")
ui_state = gr.State(None)
query = gr.Textbox(label="Query (only once)", placeholder="Enter your query and press Run")
run_btn = gr.Button("Run", variant="primary")
reset_btn = gr.Button("Reset")
with gr.Column(scale=4):
gr.Markdown(
"""
## Introduction
We introduce Re-TRAC, a recursive framework that resolves the inefficiency of isolated trials in agentic search. It not only boosts both commercial and open-source models by 15β20% over ReAct on BrowseComp but also drives SFT-only performance to new heights (30% for 4B, 53% for 30B).
</div>
<p align="center">
π€ <a href="https://github.com/microsoft/InfoAgent/tree/main/retrac">HuggingFace(comming soon)</a> |
π» <a href="https://github.com/microsoft/InfoAgent/tree/main/retrac">GitHub</a> |
π <a href="https://arxiv.org/abs/2602.02486">Paper</a> |
π <a href="https://huggingface.co/spaces/JialiangZhu/RE-TRAC">Demo</a>
</p>
## Method Overview
"""
)
gr.Image(
value="image/method.png",
label="Method Overview",
show_label=False,
)
gr.Markdown("## Main Result")
gr.Image(
value="image/main_table.png",
label="Main Result",
show_label=False,
)
run_btn.click(
run_once,
inputs=[query, ui_state],
outputs=[chatbot, ui_state, query, run_btn, status],
)
# Also allow pressing Enter (still single-run because we disable immediately)
query.submit(
run_once,
inputs=[query, ui_state],
outputs=[chatbot, ui_state, query, run_btn, status],
)
reset_btn.click(
reset,
outputs=[chatbot, ui_state, query, run_btn, status],
)
demo.queue() # required for async streaming + multiple yields
demo.launch()
|