Lenith / streaming.py
Tecnhotron
First
e78c9e1
raw
history blame
12 kB
# streaming.py - FINAL CORRECTED version with robust, state-aware processing.
import json
import re
import logging
from typing import AsyncGenerator, Callable, Optional, Iterator, Tuple
from dataclasses import dataclass
import sys, re, html
import time
from selenium.common.exceptions import NoSuchElementException, StaleElementReferenceException, TimeoutException
from selenium.webdriver.common.by import By
from selenium.webdriver.remote.webelement import WebElement
from selenium.webdriver.support.ui import WebDriverWait
from selenium.webdriver.support import expected_conditions as EC
from bs4 import BeautifulSoup, Tag
logging.basicConfig(level=logging.INFO, stream=sys.stdout, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
# --- Helper Classes (largely unchanged, but included for completeness) ---
class HtmlToMarkdownConverter:
"""
Converts a BeautifulSoup Tag element into a Markdown string.
This is designed to be called on finalized, complete HTML elements.
"""
def convert_element(self, el: Tag) -> str:
if not isinstance(el, Tag):
return ""
name = el.name
if name == 'pre':
lang_match = re.search(r'class="language-(.*?)"', str(el), re.IGNORECASE)
lang = lang_match.group(1).strip() if lang_match else ''
# .get_text() from BS4 correctly preserves newlines from the code structure
content = el.get_text().strip()
return f'```{lang}\n{content}\n```'
if name == 'p':
# Convert child tags like <strong> or <code> within the paragraph
content = ''.join(self.convert_inline(child) for child in el.contents)
return content
if name in ['ul', 'ol']:
# The site's HTML for lists already contains the bullet/number in the text.
items = [li.get_text().strip() for li in el.find_all('li', recursive=False)]
return '\n'.join(items)
if name in ['h1', 'h2', 'h3', 'h4', 'h5', 'h6']:
level = int(name[1])
return '#' * level + ' ' + el.get_text().strip()
return el.get_text()
def convert_inline(self, el):
"""Recursively converts inline elements to markdown."""
if isinstance(el, str):
return html.unescape(el)
if not isinstance(el, Tag):
return ""
content = ''.join(self.convert_inline(child) for child in el.contents)
if el.name == 'strong' or el.name == 'b':
return f'**{content}**'
if el.name == 'em' or el.name == 'i':
return f'*{content}*'
if el.name == 'code':
return f'`{content}`'
if el.name == 'a':
href = el.get('href', '')
return f'[{content}]({href})'
return content
@dataclass
class StreamConfig:
"""Configuration for streaming behavior"""
timeout: float = 300.0
retry_on_error: bool = True
max_retries: int = 3
poll_interval: float = 0.05
response_timeout: float = 900
stabilization_timeout: float = 1.0
max_inactivity: float = 10.0
stream_raw_html: bool = False
convert_html_to_markdown: bool = True
class StreamingResponseGenerator:
"""Generates Server-Sent Events (SSE) for streaming chat completions."""
def __init__(self, config: Optional[StreamConfig] = None):
self.config = config or StreamConfig()
async def create_response(
self,
completion_id: str,
created: int,
model: str,
prompt: str,
send_message_func: Callable[..., AsyncGenerator[str, None]]
) -> AsyncGenerator[str, None]:
logger.info(f"[{completion_id}] Starting streaming response generation for model '{model}'.")
first_chunk_sent = False
accumulated_content_for_logging = ""
try:
async for content_delta in send_message_func(prompt, model):
if not content_delta: continue
accumulated_content_for_logging += content_delta
delta_payload = {"content": content_delta}
if not first_chunk_sent:
delta_payload["role"] = "assistant"
first_chunk_sent = True
chunk_data = {
"id": completion_id, "object": "chat.completion.chunk", "created": created, "model": model,
"choices": [{"index": 0, "delta": delta_payload, "finish_reason": None}]
}
yield f"data: {json.dumps(chunk_data)}\n\n"
if not first_chunk_sent:
logger.warning(f"[{completion_id}] Stream ended without content. Sending empty assistant chunk.")
empty_chunk = {
"id": completion_id, "object": "chat.completion.chunk", "created": created, "model": model,
"choices": [{"index": 0, "delta": {"role": "assistant", "content": ""}, "finish_reason": None}]
}
yield f"data: {json.dumps(empty_chunk)}\n\n"
final_chunk_data = {
"id": completion_id, "object": "chat.completion.chunk", "created": created, "model": model,
"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]
}
logger.info(f"[{completion_id}] Yielding final chunk. Total content length: {len(accumulated_content_for_logging)} chars.")
yield f"data: {json.dumps(final_chunk_data)}\n\n"
logger.info(f"[{completion_id}] Yielding [DONE] signal.")
yield "data: [DONE]\n\n"
except Exception as e:
logger.error(f"[{completion_id}] Error during streaming response generation: {e}", exc_info=True)
error_payload = {"content": f"\n\nError in stream: {str(e)}"}
if not first_chunk_sent: error_payload["role"] = "assistant"
error_chunk = {
"id": completion_id, "object": "chat.completion.chunk", "created": created, "model": model,
"choices": [{"index": 0, "delta": error_payload, "finish_reason": "error" }]
}
yield f"data: {json.dumps(error_chunk)}\n\n"
yield "data: [DONE]\n\n"
# --- Main Processor with Corrected Logic ---
class StreamProcessor:
"""Processes streaming content from Selenium, with stabilization and content conversion."""
def __init__(self, config: Optional[StreamConfig] = None):
cfg = config or StreamConfig()
self.poll_interval = cfg.poll_interval
self.response_timeout = cfg.response_timeout
self.stabilization_timeout = cfg.stabilization_timeout
self.max_inactivity = cfg.max_inactivity
self.request_id_for_log = "StreamProc"
self.markdown_converter = HtmlToMarkdownConverter()
def set_request_id(self, request_id: str):
self.request_id_for_log = request_id
def get_processed_text_stream(self, driver, element_locator: Tuple[str, str]) -> Iterator[str]:
html_stream = self._poll_element_content_stream(driver, element_locator)
return self._convert_html_stream_to_markdown_deltas(html_stream)
def _convert_html_stream_to_markdown_deltas(self, html_iterator: Iterator[str]) -> Iterator[str]:
"""
Statefully converts a stream of HTML snapshots to Markdown deltas.
It identifies "finalized" vs "active" blocks to avoid streaming
volatile, incomplete code blocks, preventing corruption.
"""
last_yielded_markdown = ""
last_html_snapshot = ""
for full_html_snapshot in html_iterator:
last_html_snapshot = full_html_snapshot
soup = BeautifulSoup(f"<ol>{full_html_snapshot}</ol>", 'lxml')
# Find all content blocks in the last message bubble.
# The structure is <ol> -> <div> (bubble) -> ... -> <div class="prose"> -> <p/pre/ul>
all_prose_divs = soup.select('div.prose')
if not all_prose_divs:
continue
content_elements = [div.contents[0] for div in all_prose_divs if div.contents]
finalized_elements = content_elements
# Check if the very last element is an incomplete code block.
# If so, we don't process it in this pass, we wait for it to be finalized.
if content_elements and content_elements[-1].name == 'pre':
finalized_elements = content_elements[:-1]
# Convert all finalized elements to markdown
md_pieces = [self.markdown_converter.convert_element(el) for el in finalized_elements]
current_safe_markdown = "\n\n".join(md_pieces)
# Yield the delta if it's a simple append
if current_safe_markdown != last_yielded_markdown:
if current_safe_markdown.startswith(last_yielded_markdown):
delta = current_safe_markdown[len(last_yielded_markdown):]
if delta:
yield delta.lstrip('\n')
last_yielded_markdown = current_safe_markdown
# After the loop, the stream has finished. Process the very last snapshot in full.
final_soup = BeautifulSoup(f"<ol>{last_html_snapshot}</ol>", 'lxml')
all_prose_divs = final_soup.select('div.prose')
content_elements = [div.contents[0] for div in all_prose_divs if div.contents]
final_md_pieces = [self.markdown_converter.convert_element(el) for el in content_elements]
final_markdown = "\n\n".join(final_md_pieces)
if final_markdown.startswith(last_yielded_markdown):
final_delta = final_markdown[len(last_yielded_markdown):]
if final_delta:
yield final_delta.lstrip('\n')
def _poll_element_content_stream(self, driver, element_locator: Tuple[str, str]) -> Iterator[str]:
log_prefix = f"[{self.request_id_for_log}/PollStream]"
start_time = time.time()
last_content = ""
last_change_time = time.time()
while time.time() - start_time < self.response_timeout:
loop_start_time = time.time()
try:
element = WebDriverWait(driver, self.poll_interval * 2).until(EC.presence_of_element_located(element_locator))
current_content = element.get_attribute('innerHTML')
if current_content != last_content:
yield current_content
last_content = current_content
last_change_time = time.time()
elif time.time() - last_change_time > self.max_inactivity:
logger.info(f"{log_prefix} Content stable for {self.max_inactivity:.2f}s. Ending poll.")
return
except (TimeoutException, StaleElementReferenceException, NoSuchElementException):
logger.debug(f"{log_prefix} Element not present or stale. Continuing poll.")
except Exception as e:
logger.error(f"{log_prefix} Unexpected error polling: {e}", exc_info=True)
return
sleep_duration = self.poll_interval - (time.time() - loop_start_time)
if sleep_duration > 0: time.sleep(sleep_duration)
logger.warning(f"{log_prefix} Polling finished due to max wait ({self.response_timeout:.2f}s).")
async def create_streaming_response(
completion_id: str,
created: int,
model: str,
prompt: str,
send_message_func: Callable[..., AsyncGenerator[str, None]],
stream_config: Optional[StreamConfig] = None
) -> AsyncGenerator[str, None]:
generator = StreamingResponseGenerator(config=stream_config)
async for chunk in generator.create_response(completion_id, created, model, prompt, send_message_func):
yield chunk
__all__ = ['create_streaming_response', 'StreamingResponseGenerator', 'StreamConfig', 'StreamProcessor', 'HtmlToMarkdownConverter']