Pass TestSearch benchmark consistently (Add browse_website TOKENS_TO_…

…TRIGGER_SUMMARY) (Significant-Gravitas#5092)

* Added SUMMARIZATION_TRIGGER_LENGTH
browse_website won't summarize content that's shorter
than SUMMARIZATION_TRIGGER_LENGTH.
It defaults to 250 characters, which is approximately 50 tokens.

* Refactor BrowserOptions

* Use tokens instead of length
to trigger summarization

* Bugfix

* fix: Always return links even if not summarizing
feat: Increase the number of links returned from 5 to 20

---------

Co-authored-by: lc0rp <[email protected]>
Co-authored-by: James Collins <[email protected]>

autogpt/commands/web_selenium.py

@@ -2,20 +2,23 @@
 
 from __future__ import annotations
 
+from autogpt.llm.utils.token_counter import count_string_tokens
+
 COMMAND_CATEGORY = "web_browse"
 COMMAND_CATEGORY_TITLE = "Web Browsing"
 
 import logging
 from pathlib import Path
 from sys import platform
-from typing import Optional, Type
+from typing import Optional
 
 from bs4 import BeautifulSoup
 from selenium.common.exceptions import WebDriverException
 from selenium.webdriver.chrome.options import Options as ChromeOptions
 from selenium.webdriver.chrome.service import Service as ChromeDriverService
 from selenium.webdriver.chrome.webdriver import WebDriver as ChromeDriver
 from selenium.webdriver.common.by import By
+from selenium.webdriver.common.options import ArgOptions as BrowserOptions
 from selenium.webdriver.edge.options import Options as EdgeOptions
 from selenium.webdriver.edge.service import Service as EdgeDriverService
 from selenium.webdriver.edge.webdriver import WebDriver as EdgeDriver
@@ -38,9 +41,9 @@
 from autogpt.processing.html import extract_hyperlinks, format_hyperlinks
 from autogpt.url_utils.validators import validate_url
 
-BrowserOptions = ChromeOptions | EdgeOptions | FirefoxOptions | SafariOptions
-
 FILE_DIR = Path(__file__).parent.parent
+TOKENS_TO_TRIGGER_SUMMARY = 50
+LINKS_TO_RETURN = 20
 
 
 @command(
@@ -64,25 +67,30 @@ def browse_website(url: str, question: str, agent: Agent) -> str:
         question (str): The question asked by the user
 
     Returns:
-        Tuple[str, WebDriver]: The answer and links to the user and the webdriver
+        str: The answer and links to the user and the webdriver
     """
+    driver = None
     try:
         driver, text = scrape_text_with_selenium(url, agent)
+        add_header(driver)
+        if TOKENS_TO_TRIGGER_SUMMARY < count_string_tokens(text, agent.llm.name):
+            text = summarize_memorize_webpage(url, text, question, agent, driver)
+
+        links = scrape_links_with_selenium(driver, url)
+
+        # Limit links to LINKS_TO_RETURN
+        if len(links) > LINKS_TO_RETURN:
+            links = links[:LINKS_TO_RETURN]
+
+        return f"Answer gathered from website: {text}\n\nLinks: {links}"
     except WebDriverException as e:
         # These errors are often quite long and include lots of context.
         # Just grab the first line.
         msg = e.msg.split("\n")[0]
         return f"Error: {msg}"
-
-    add_header(driver)
-    summary = summarize_memorize_webpage(url, text, question, agent, driver)
-    links = scrape_links_with_selenium(driver, url)
-
-    # Limit links to 5
-    if len(links) > 5:
-        links = links[:5]
-    close_browser(driver)
-    return f"Answer gathered from website: {summary}\n\nLinks: {links}"
+    finally:
+        if driver:
+            close_browser(driver)
 
 
 def scrape_text_with_selenium(url: str, agent: Agent) -> tuple[WebDriver, str]:
@@ -96,7 +104,7 @@ def scrape_text_with_selenium(url: str, agent: Agent) -> tuple[WebDriver, str]:
     """
     logging.getLogger("selenium").setLevel(logging.CRITICAL)
 
-    options_available: dict[str, Type[BrowserOptions]] = {
+    options_available: dict[str, BrowserOptions] = {
         "chrome": ChromeOptions,
         "edge": EdgeOptions,
         "firefox": FirefoxOptions,