exposed flask server to invoke agent

Author: nischalj10

README.md

@@ -88,6 +88,17 @@ primaryClass={cs.AI},
 url={https://arxiv.org/abs/2408.07199},
 }
 
+```
+```
+@inproceedings{yao2022webshop,
+  bibtex_show = {true},
+  title = {WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents},
+  author = {Yao, Shunyu and Chen, Howard and Yang, John and Narasimhan, Karthik},
+  booktitle = {ArXiv},
+  year = {preprint},
+  html = {https://arxiv.org/abs/2207.01206},
+  tag = {NLP}
+}
 ```
 
 ```

agentq/__main__.py

@@ -1,5 +1,7 @@
 import asyncio
 
+from playwright.async_api import Page
+
 from agentq.core.agent.agentq import AgentQ
 from agentq.core.agent.agentq_actor import AgentQActor
 from agentq.core.agent.agentq_critic import AgentQCritic
@@ -8,17 +10,36 @@
 from agentq.core.models.models import State
 from agentq.core.orchestrator.orchestrator import Orchestrator
 
+state_to_agent_map = {
+    State.PLAN: PlannerAgent(),
+    State.BROWSE: BrowserNavAgent(),
+    State.AGENTQ_BASE: AgentQ(),
+    State.AGENTQ_ACTOR: AgentQActor(),
+    State.AGENTQ_CRITIC: AgentQCritic(),
+}
+
+
+async def run_agent(command):
+    orchestrator = Orchestrator(state_to_agent_map=state_to_agent_map, eval_mode=True)
+    await orchestrator.start()
+    page: Page = await orchestrator.playwright_manager.get_current_page()
+    await page.set_extra_http_headers({"User-Agent": "AgentQ-Bot"})
+    await page.goto("https://686b-49-205-35-248.ngrok-free.app/abc", timeout=30000)
+    result = await orchestrator.execute_command(command)
+    return result
 
-async def main():
-    # Define state machine
-    state_to_agent_map = {
-        State.PLAN: PlannerAgent(),
-        State.BROWSE: BrowserNavAgent(),
-        State.AGENTQ_BASE: AgentQ(),
-        State.AGENTQ_ACTOR: AgentQActor(),
-        State.AGENTQ_CRITIC: AgentQCritic(),
-    }
 
+def run_agent_sync(command):
+    if asyncio.get_event_loop().is_closed():
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+    else:
+        loop = asyncio.get_event_loop()
+
+    return loop.run_until_complete(run_agent(command))
+
+
+async def main():
     orchestrator = Orchestrator(state_to_agent_map=state_to_agent_map)
     await orchestrator.start()
 

agentq/core/agent/agentq.py

@@ -29,7 +29,7 @@ def __modify_system_prompt(self, ltm):
         system_prompt: str = LLM_PROMPTS["AGENTQ_BASE_PROMPT"]
 
         substitutions = {
-            "basic_user_information": ltm if ltm is not None else "",
+            "task_information": ltm if ltm is not None else "",
         }
 
         # Use safe_substitute to avoid KeyError

agentq/core/orchestrator/orchestrator.py

@@ -49,6 +49,7 @@ def __init__(
         self.eval_mode = eval_mode
         self.shutdown_event = asyncio.Event()
         self.session_id = str(uuid.uuid4())
+        self.memory = None
 
     async def start(self):
         print("Starting orchestrator")
@@ -95,9 +96,12 @@ async def execute_command(self, command: str):
                 current_tasks_for_eval=None,
                 sorted_tasks=None,
             )
+            # Get the current event loop
+            loop = asyncio.get_event_loop()
+
             print(f"Executing command {self.memory.objective}")
             while self.memory.current_state != State.COMPLETED:
-                await self._handle_state()
+                await loop.create_task(self._handle_state())
             self._print_final_response()
 
             if self.eval_mode:

agentq/core/prompts/prompts.py

@@ -185,6 +185,7 @@ class BrowserNavOutput(BaseModel):
  4. If the task requires multiple informations, all of them are equally important and should be gathered before terminating the task. You will strive to meet all the requirements of the task.
  5. If one plan fails, you MUST revise the plan and try a different approach. You will NOT terminate a task untill you are absolutely convinced that the task is impossible to accomplish.
  6. Think critically if the task has been actually been achieved before doing the final termination.
+ 7. Make sure to take into account task sepcific information.
 
  ## Web Navigation guidelines ##
  1. Based on the actions you output, web navigation will be done, which may include logging into websites and interacting with any web content
@@ -245,7 +246,7 @@ class BrowserNavOutput(BaseModel):
 
 Notice above how there is confirmation after each step and how interaction (e.g. setting source and destination) with each element is a separate step. Follow same pattern.
 
-Some basic information about the user: \n $basic_user_information
+Some task sepcific information that you MUST take into account: \n $task_information
 
  ## SOME VERY IMPORTANT POINTS TO ALWAYS REMEMBER ##
  1. NEVER ASK WHAT TO DO NEXT or HOW would you like to proceed to the user.

agentq/core/skills/open_url.py

@@ -40,6 +40,9 @@ async def openurl(
         try:
             await browser_manager.take_screenshots(f"{function_name}_start", page)
 
+            # set extra headers for bypassing ngrok
+            await page.set_extra_http_headers({"User-Agent": "AgentQ-Sentient"})
+
             # Use a longer timeout for navigation
             await page.goto(
                 url, timeout=max(30000, timeout * 1000), wait_until="domcontentloaded"

agentq/core/web_driver/playwright.py

@@ -270,14 +270,11 @@ async def get_current_page(self) -> Page:
                 page: Page = await browser.new_page()  # type: ignore
                 # await stealth_async(page)  # Apply stealth to the new page
                 return page
-        except Exception as e:
-            logger.warn(f"Browser context was closed. Creating a new one. {e}")
         except Exception as e:
             logger.warn(f"Browser context was closed. Creating a new one. {e}")
             PlaywrightManager._browser_context = None
-            _browser: BrowserContext = await self.get_browser_context()  # type: ignore
-            page: Union[Page, None] = await self.get_current_page()
-            return page
+            await self.ensure_browser_context()
+            return await self.get_current_page()
 
     async def close_all_tabs(self, keep_first_tab: bool = True):
         """

agentq/user_preferences/user_preferences.txt

@@ -1,26 +1,7 @@
-Personal Info:
-First Name: Nischal
-Last Name: Jain
-Date of birth: 
-Pronouns: He/Him
-
-Email: 
-Expeted Salary: 5000000
-Occupation: Data Engineer
-Address: Indiranagr, Bengaluru
-Phone Number: 
-LinkedIn - https://linkedin.com/in/nischalj10
-
-
-Here are some of my preferences:
-Shopping Preferences: www.amazon.com
-Favorite news source: www.thehindu.com
-Favorite flight booking site to use with every flight related query: https://www.google.com/travel/flightsno
-Resume File to Upload Path = /Users/namanjain/Downloads/DummyResume.pdf
-My best project - checkout headless-ollama on google
-
-
-I am comfortable commuting to office.
-I have 4 years of experience in TypeScript, Node Js, Nextjs.
-I have work authorization for working in the USA.
-I have experience in cloud infrastructure, data warheouses.
+1. Your job is to find the relevant product asked by the user on a locally hosted ecommerce website https://686b-49-205-35-248.ngrok-free.app 
+2. Your search query should be small and should only have important keywords. In general, you should try to checkout the first couple of results shown in the search results. 
+3. Remember, a lot of custom filters like size, color, flavour, quantity etc are only available on the product page and the product title in search page may not exactly fit requirements but can loosely relate to the search query - so checkout first couple of search results that match most closely. 
+    Eg - you are asked to  buy strawberry biscuits from birtannia - but when you search britannia strawberry biscuits, you get search reuslt with chocloate flavoured britannia buiscuits. You should still click on it as there could be an option on the product page to change falvour from chocolate to strawberry because the brand is same and the product is same "braitannia and buiscuits"
+4. If you do not find the relevant product in first couple of search results pages, you should go back to the search page and try to search with a different query.
+5. Make sure to pay attention to all the attributes like size/ color/ quantity etc. mentioned in the user's query and select appropriate attributes on the product details page before buying it. 
+6. Ultimately - your task will only end when you click on the "Buy Now" button of the right product.

poetry.lock

@@ -27,6 +27,7 @@ tabulate = "^0.9.0"
 nltk = "^3.9.1"
 langsmith = "^0.1.104"
 instructor = "^1.4.0"
+flask = "^3.0.3"
 numpy = "^2.1.0"