peijun1's picture
Deploy AI Studio Proxy API to Hugging Face Spaces
a5784e9
Raw
History Blame Contribute Delete
22.5 kB
import asyncio
import json
import queue
import re
import time
from typing import Any, AsyncGenerator, Callable, List, Optional, Tuple
from config.settings import FUNCTION_CALLING_DEBUG
from logging_utils import set_request_id
# [REFAC-01] Structural Boundary Pattern
TOOL_STRUCTURE_PATTERN = re.compile(
r"(?:^|\n)\s*(?:```[a-zA-Z0-9]*\s*)?<[a-zA-Z0-9_\-]+(?:\s|>)"
)
async def use_stream_response(
req_id: str,
timeout: float = 5.0,
silence_threshold: float = 60.0,
page: Any = None,
check_client_disconnected: Optional[Callable] = None,
stream_start_time: float = 0.0,
enable_silence_detection: bool = True,
) -> AsyncGenerator[Any, None]:
"""Enhanced stream response handler with UI-based generation active checks."""
from api_utils.server_state import state
STREAM_QUEUE = state.STREAM_QUEUE
logger = state.logger
from browser_utils.page_controller import PageController
from config import (
CHAT_SESSION_CONTENT_SELECTOR,
LAST_CHAT_TURN_SELECTOR,
SCROLL_CONTAINER_SELECTOR,
UI_GENERATION_WAIT_TIMEOUT_MS,
)
from config.global_state import GlobalState
from models import (
ClientDisconnectedError,
QuotaExceededError,
UpstreamError,
)
set_request_id(req_id)
if STREAM_QUEUE is None:
logger.warning(f"[{req_id}] STREAM_QUEUE is None, cannot use stream response")
return
if stream_start_time == 0.0:
stream_start_time = time.time() - 10.0
accumulated_body = ""
accumulated_reason_len = 0
total_reason_processed = 0
total_body_processed = 0
boundary_transitions = 0
boundary_buffer = ""
acc_reason_state = ""
acc_body_state = ""
force_body_mode = False
split_index = -1
empty_count = 0
initial_wait_limit = int(timeout * 10)
silence_wait_limit = int(silence_threshold * 10)
max_empty_retries = max(silence_wait_limit, initial_wait_limit)
hard_timeout_limit = int(timeout * 10 * 3)
_data_received = False
has_content = False
has_seen_functions = False
received_items_count = 0
stale_done_ignored = False
last_ui_check_time = 0
ui_check_interval = int(UI_GENERATION_WAIT_TIMEOUT_MS / 100)
if ui_check_interval <= 0:
ui_check_interval = 1
last_packet_time = time.time()
min_items_before_silence_check = 10
async def check_ui_generation_active():
if not page:
return False
try:
stop_button = page.locator('button[aria-label="Stop generating"]')
if await stop_button.is_visible(timeout=1000):
return True
from config.selectors import SUBMIT_BUTTON_SELECTOR
submit_button = page.locator(SUBMIT_BUTTON_SELECTOR)
if await submit_button.count() > 0:
try:
if await submit_button.first.is_disabled(timeout=2000):
return True
except Exception:
return False
return False
except Exception:
return False
try:
while True:
if (
GlobalState.CURRENT_STREAM_REQ_ID
and GlobalState.CURRENT_STREAM_REQ_ID != req_id
):
logger.warning(f"[{req_id}] Zombie Stream detected. Terminating.")
yield {
"done": True,
"reason": "zombie_stream_aborted",
"body": "",
"function": [],
}
return
if page:
try:
await page.evaluate(
"""([scrollSel, contentSel, lastTurnSel]) => {
const scrollContainer = document.querySelector(scrollSel);
if (scrollContainer) scrollContainer.scrollTop = scrollContainer.scrollHeight;
const sessionContent = document.querySelector(contentSel);
if (sessionContent) sessionContent.scrollTop = sessionContent.scrollHeight;
const lastTurn = document.querySelector(lastTurnSel);
if (lastTurn) lastTurn.scrollIntoView({behavior: "instant", block: "end"});
window.scrollTo(0, document.body.scrollHeight);
}""",
[
SCROLL_CONTAINER_SELECTOR,
CHAT_SESSION_CONTENT_SELECTOR,
LAST_CHAT_TURN_SELECTOR,
],
)
except Exception:
pass
if GlobalState.IS_QUOTA_EXCEEDED and not GlobalState.IS_RECOVERING:
logger.warning(f"[{req_id}] Quota detected. Pausing...")
try:
start_wait = time.time()
while time.time() - start_wait < 2.0:
if GlobalState.IS_RECOVERING:
break
await asyncio.sleep(0.2)
except Exception:
pass
if GlobalState.IS_RECOVERING:
logger.info(f"[{req_id}] 🔄 Recovery mode detected. Holding...")
elif not GlobalState.IS_QUOTA_EXCEEDED:
logger.info(f"[{req_id}] ✅ Recovery completed. Resuming...")
else:
logger.warning(f"[{req_id}] ⛔ Quota exceeded, waiting...")
await asyncio.sleep(1.0)
continue
if GlobalState.IS_SHUTTING_DOWN.is_set():
logger.warning(f"[{req_id}] 🛑 Global Shutdown. Aborting.")
yield {
"done": True,
"reason": "global_shutdown",
"body": "",
"function": [],
}
return
try:
data = STREAM_QUEUE.get_nowait()
if data is None:
logger.info(f"[{req_id}] 🔴 Received termination signal.")
break
if isinstance(data, dict) and data.get("done") is True:
logger.info(f"[{req_id}] ✅ Explicit DONE received.")
yield data
break
empty_count = 0
_data_received = True
received_items_count += 1
last_packet_time = time.time()
actual_data = data
if isinstance(data, str):
try:
parsed_wrapper = json.loads(data)
if (
isinstance(parsed_wrapper, dict)
and "ts" in parsed_wrapper
and "data" in parsed_wrapper
):
if parsed_wrapper["ts"] < stream_start_time:
logger.warning(f"[{req_id}] 🗑️ Stale data ignored.")
continue
actual_data = parsed_wrapper["data"]
else:
actual_data = parsed_wrapper
except json.JSONDecodeError:
pass
if isinstance(actual_data, dict):
if actual_data.get("error"):
status = actual_data.get("status", 500)
message = actual_data.get("message", "Unknown error")
if status == 429 or "quota" in message.lower():
raise QuotaExceededError(
f"AI Studio quota exceeded: {message}", req_id=req_id
)
else:
raise UpstreamError(
f"AI Studio error: {message}",
status_code=status,
)
parsed_data = actual_data
p_reason = str(parsed_data.get("reason", ""))
p_body = str(parsed_data.get("body", ""))
if (
p_reason
and acc_reason_state
and p_reason.startswith(acc_reason_state)
):
new_reason_delta = p_reason[len(acc_reason_state) :]
acc_reason_state = p_reason
else:
acc_reason_state += p_reason
new_reason_delta = p_reason
if p_body and acc_body_state and p_body.startswith(acc_body_state):
acc_body_state = p_body
else:
acc_body_state += p_body
if force_body_mode:
parsed_data["reason"] = acc_reason_state[:split_index]
parsed_data["body"] = (
acc_body_state + acc_reason_state[split_index:]
)
else:
text_to_check = boundary_buffer + new_reason_delta
match = TOOL_STRUCTURE_PATTERN.search(text_to_check)
if match:
offset = len(acc_reason_state) - len(text_to_check)
split_index = offset + match.start()
force_body_mode = True
boundary_transitions += 1
parsed_data["reason"] = acc_reason_state[:split_index]
parsed_data["body"] = (
acc_body_state + acc_reason_state[split_index:]
)
logger.info(f"[{req_id}] ✂️ Boundary Split Applied.")
else:
parsed_data["reason"] = acc_reason_state
parsed_data["body"] = acc_body_state
boundary_buffer = (boundary_buffer + new_reason_delta)[
-100:
]
accumulated_body += str(parsed_data.get("body", ""))
accumulated_reason_len += len(str(parsed_data.get("reason", "")))
total_body_processed += len(str(parsed_data.get("body", "")))
total_reason_processed += len(str(parsed_data.get("reason", "")))
if parsed_data.get("body") or parsed_data.get("reason"):
has_content = True
if parsed_data.get("function"):
has_seen_functions = True
# Track if any function call has empty arguments (potential parse failure)
for fc in parsed_data.get("function", []):
fc_params = fc.get("params") or fc.get("arguments") or {}
if not fc_params:
if FUNCTION_CALLING_DEBUG:
logger.warning(
f"[{req_id}] ⚠️ Wire format returned '{fc.get('name')}' with empty args - will try DOM fallback"
)
has_seen_functions = False # Force DOM fallback
break
if parsed_data.get("done") is True:
if GlobalState.IS_QUOTA_EXCEEDED or GlobalState.IS_RECOVERING:
logger.info(
f"[{req_id}] 🛡️ Quota/Recovery active: Holding stream open."
)
continue
just_rotated = (
time.time() - GlobalState.LAST_ROTATION_TIMESTAMP < 15.0
)
recently_recovered = (
time.time() - GlobalState.LAST_ROTATION_TIMESTAMP < 30.0
)
if (
not has_content
and received_items_count == 1
and not stale_done_ignored
and not GlobalState.IS_QUOTA_EXCEEDED
and (just_rotated or recently_recovered)
):
logger.info(
f"[{req_id}] 🔄 Post-rotation empty DONE detected. Ignoring."
)
stale_done_ignored = True
continue
if (
parsed_data.get("done") is True
and not has_seen_functions
and page
):
# Retry loop for DOM function call detection - UI elements may not render immediately
# Similar to body text retry loop below, but shorter timeout for function calls
dom_functions = []
dom_text = ""
max_fc_retries = 10 # 10 retries * 0.3s = 3 seconds max wait
for fc_retry in range(max_fc_retries):
(
dom_functions,
dom_text,
) = await detect_function_calls_from_dom(
page, req_id, logger
)
if dom_functions:
if FUNCTION_CALLING_DEBUG:
logger.info(
f"[{req_id}] ✅ DOM captured function calls after {fc_retry + 1} attempts"
)
break
# Only retry if we haven't found functions and body is also empty
# (indicates potential race condition with UI rendering)
if accumulated_body:
break # We have body text, no need to wait for functions
await asyncio.sleep(0.3)
if dom_functions:
parsed_data["function"] = dom_functions
has_seen_functions = True
# If we have DOM text and accumulated body is empty, inject it to final chunk
if dom_text and not accumulated_body:
parsed_data["body"] = dom_text
accumulated_body = dom_text
yield parsed_data
if parsed_data.get("done") is True:
if (
accumulated_reason_len > 0
and len(accumulated_body) == 0
and not has_seen_functions
):
logger.info(
f"[{req_id}] ⚠️ Thinking-Only response detected. Waiting for DOM..."
)
try:
if page:
pc = PageController(page, logger, req_id)
for _ in range(20):
await asyncio.sleep(0.5)
dom_text = (
await pc.get_body_text_only_from_dom()
)
if dom_text and len(dom_text.strip()) > 0:
logger.info(
f"[{req_id}] ✅ DOM captured body: {len(dom_text)} chars"
)
yield {
"body": dom_text,
"reason": "",
"done": False,
}
break
except Exception as e:
logger.error(f"[{req_id}] DOM Wait Error: {e}")
break
else:
stale_done_ignored = False
continue
except (queue.Empty, asyncio.QueueEmpty):
empty_count += 1
if (
enable_silence_detection
and received_items_count >= min_items_before_silence_check
and time.time() - last_packet_time > silence_threshold
):
logger.info(f"[{req_id}] 🔇 Stream silence detected.")
yield {
"done": True,
"reason": "silence_detected",
"body": "",
"function": [],
}
return
if empty_count % 50 == 0:
logger.info(
f"[{req_id}] Waiting for data... ({empty_count}/{max_empty_retries})"
)
if empty_count >= max_empty_retries:
if GlobalState.IS_RECOVERING:
empty_count = 0
continue
if (
await check_ui_generation_active()
and empty_count < hard_timeout_limit
):
logger.warning(f"[{req_id}] Timeout but UI active. Snoozing...")
empty_count = max(0, empty_count - int(max_empty_retries * 0.5))
continue
elif empty_count >= hard_timeout_limit:
logger.error(f"[{req_id}] HARD TIMEOUT REACHED!")
yield {
"done": True,
"reason": "hard_timeout",
"body": "",
"function": [],
}
return
yield {
"done": True,
"reason": "internal_timeout",
"body": "",
"function": [],
}
return
if check_client_disconnected:
try:
check_client_disconnected(f"Stream Queue Wait ({req_id})")
except ClientDisconnectedError:
raise
if received_items_count == 0 and empty_count >= initial_wait_limit:
logger.error(f"[{req_id}] Stream has no data (TTFB Timeout).")
try:
from api_utils.server_state import state
page_instance = state.page_instance
if page_instance:
await page_instance.reload()
except Exception:
pass
yield {
"done": True,
"reason": "ttfb_timeout",
"body": "",
"function": [],
}
return
if empty_count - last_ui_check_time >= ui_check_interval:
if await check_ui_generation_active():
logger.info(f"[{req_id}] UI detected still generating...")
last_ui_check_time = empty_count
await asyncio.sleep(0.1)
continue
except asyncio.CancelledError:
raise
except Exception as e:
if isinstance(e, ClientDisconnectedError):
raise e
logger.error(f"[{req_id}] Error in stream generator: {e}", exc_info=True)
raise
finally:
logger.info(
f"[{req_id}] Stream response completed. Items: {received_items_count}"
)
await clear_stream_queue()
async def clear_stream_queue():
import queue
from api_utils.server_state import state
STREAM_QUEUE = state.STREAM_QUEUE
logger = state.logger
if STREAM_QUEUE is None:
return
cleared_count = 0
while True:
try:
await asyncio.to_thread(STREAM_QUEUE.get_nowait)
cleared_count += 1
except queue.Empty:
break
except Exception:
break
if cleared_count > 0:
logger.info(f"Stream queue cleared. Items: {cleared_count}")
async def detect_function_calls_from_dom(
page: Any,
req_id: str,
logger: Any,
) -> Tuple[List[dict], str]:
"""Fallback function call detection using DOM parsing.
This is used when the network interceptor doesn't capture function calls
(e.g., due to timing issues or format changes).
Args:
page: Playwright page instance.
req_id: Request ID for logging.
logger: Logger instance.
Returns:
Tuple of (List of function call dicts, text content).
"""
if not page:
return [], ""
try:
from api_utils.utils_ext.function_call_response_parser import (
FunctionCallResponseParser,
)
parser = FunctionCallResponseParser(page, logger, req_id)
result = await parser.parse_function_calls()
function_calls: List[dict] = []
if result.has_function_calls and result.function_calls:
# Convert ParsedFunctionCall objects to dict format expected by stream
for fc in result.function_calls:
function_calls.append({"name": fc.name, "params": fc.arguments})
if FUNCTION_CALLING_DEBUG:
logger.info(
f"[{req_id}] DOM fallback detected {len(function_calls)} function call(s)"
)
return function_calls, result.text_content
except Exception as e:
if FUNCTION_CALLING_DEBUG:
logger.debug(f"[{req_id}] DOM function call detection failed: {e}")
return [], ""