stillerman's picture
stillerman HF Staff
docs minus assets
1e310b7
from typing import List, Tuple, Dict, Optional
import sqlite3
import json
import litellm
import re
import asyncio
import argparse
from functools import lru_cache
class SQLiteDB:
def __init__(self, db_path: str):
"""Initialize the database with path to SQLite database"""
self.db_path = db_path
self.conn = sqlite3.connect(db_path)
self.conn.row_factory = sqlite3.Row
self.cursor = self.conn.cursor()
self._article_count = self._get_article_count()
print(f"Connected to SQLite database with {self._article_count} articles")
def _get_article_count(self):
self.cursor.execute("SELECT COUNT(*) FROM core_articles")
return self.cursor.fetchone()[0]
@lru_cache(maxsize=8192)
def get_article_with_links(self, article_title: str) -> Tuple[str, List[str]]:
self.cursor.execute(
"SELECT title, links_json FROM core_articles WHERE title = ?",
(article_title,),
)
article = self.cursor.fetchone()
if not article:
return None, []
links = json.loads(article["links_json"])
return article["title"], links
class Player:
def __init__(self, name: str):
self.name = name
async def get_move(self, game_state: List[Dict]) -> Tuple[str, Dict]:
print("Link choices:")
for i, link in enumerate(game_state[-1]["links"]):
print(f"{i}: {link}")
idx = int(input(f"Enter the index of the link you want to select: "))
return game_state[-1]["links"][idx], {
"message": f"{self.name} selected link #{i}"
} # select the first link
class AgentPlayer(Player):
def __init__(
self,
model: str,
api_base: str,
verbose: bool = True,
max_links=None,
max_tries=10,
target_article = None,
seed = None
):
super().__init__(model)
self.model = model
self.api_base = api_base
self.verbose = verbose
self.max_links = max_links
self.max_tries = max_tries
self.target_article = target_article
self.seed = seed
async def get_move(self, game_state: List[Dict]) -> Tuple[str, Dict]:
prompt = self.construct_prompt(game_state)
conversation = [
{"role": "user", "content": prompt}
]
for try_number in range(self.max_tries):
response = await litellm.acompletion(
model=self.model,
api_base=self.api_base,
messages=conversation,
seed=self.seed
)
response = response.choices[0].message.content
conversation.append({"role": "assistant", "content": response})
answer, message = self._attempt_to_extract_answer(response, maximum_answer=len(game_state[-1]["links"]))
# there was a problem with the answer so give the model another chance
if answer == -1:
conversation.append({"role": "user", "content": message})
continue
assert answer >= 1 and answer <= len(game_state[-1]["links"]), f"Answer {answer} is out of range"
# we found an answer so we can return it
return game_state[-1]["links"][answer-1], {"tries": try_number, "conversation": conversation}
# we tried the max number of times and still didn't find an answer
return -1, {"tries": self.max_tries, "conversation": conversation}
def construct_prompt(self, game_state: List[Dict]) -> str:
current = game_state[-1]["article"]
target = self.target_article
available_links = game_state[-1]["links"]
formatted_links = "\n".join([f"{i+1}. {link}" for i, link in enumerate(available_links)])
path_so_far = [step["article"] for step in game_state]
try:
formatted_path = ' -> '.join(path_so_far)
except Exception as e:
print(f"Error formatting path: {e}")
print(game_state)
print("Path so far: ", path_so_far)
raise e
return f"""You are playing WikiRun, trying to navigate from one Wikipedia article to another using only links.
IMPORTANT: You MUST put your final answer in <answer>NUMBER</answer> tags, where NUMBER is the link number.
For example, if you want to choose link 3, output <answer>3</answer>.
Current article: {current}
Target article: {target}
Available links (numbered):
{formatted_links}
Your path so far: {formatted_path}
Think about which link is most likely to lead you toward the target article.
First, analyze each link briefly and how it connects to your goal, then select the most promising one.
Remember to format your final answer by explicitly writing out the xml number tags like this: <answer>NUMBER</answer>
"""
def _attempt_to_extract_answer(self, response: str, maximum_answer: Optional[int] = None) -> Tuple[int, str]:
'returns -1 and a message if no answer is found'
# Extract choice using format <answer>N</answer>
choice_match = re.search(r"<answer>(\d+)</answer>", response)
if choice_match is None:
return -1, f"No answer found in response. Please respond with a number between 1 and {maximum_answer} in <answer>NUMBER</answer> tags."
# check if there are multiple answers
multiple_answers = re.findall(r"<answer>(\d+)</answer>", response)
if len(multiple_answers) > 1:
return -1, "Multiple answers found in response. Please respond with just one."
answer = choice_match.group(1)
# try to convert to int
try:
answer = int(answer)
except ValueError:
return -1, f"You answered with {answer} but it could not be converted to an integer. Please respond with a number between 1 and {maximum_answer}."
# check if the answer is too high or too low
if answer > maximum_answer or answer < 1:
return -1, f"You answered with {answer} but you have to select a number between 1 and {maximum_answer}."
return answer, "" # we found an answer so we don't need to return a message
class Game:
def __init__(
self,
start_article: str,
target_article: str,
db: SQLiteDB,
max_allowed_steps: int,
player: Player,
verbose: bool = True,
):
self.start_article = start_article
self.target_article = target_article
self.db = db
self.max_allowed_steps = max_allowed_steps
self.steps = []
self.steps_taken = 0
self.player = player
self.verbose = verbose
# Ensure the player knows the target article
if isinstance(self.player, AgentPlayer):
self.player.target_article = self.target_article
async def run(self):
if self.verbose:
print(f"Starting game from {self.start_article} to {self.target_article}")
# get the start article
_, links = self.db.get_article_with_links(self.start_article)
self.steps.append(
{
"type": "start",
"article": self.start_article,
"links": links,
"metadata": {"message": "Game started"},
}
)
# while the current article is not the target article and the number of steps taken is less than the max allowed steps
while self.steps_taken < self.max_allowed_steps:
self.steps_taken += 1
# Await the async player move
player_move, metadata = await self.player.get_move(self.steps)
# player couldn't select a valid link
if player_move == -1:
self.steps.append(
{"type": "lose", "article": player_move, "metadata": metadata}
)
break
if self.verbose:
print(f" -> Step {self.steps_taken}: {player_move}")
# input("Press Enter to continue...")
# if we found it its over
if player_move == self.target_article:
self.steps.append(
{"type": "win", "article": player_move, "metadata": metadata}
)
break
# if not lets get the next article
_, links = self.db.get_article_with_links(player_move)
if len(links) == 0:
self.steps.append(
{"type": "lose", "article": player_move, "metadata": metadata}
)
break
self.steps.append(
{
"type": "move",
"article": player_move,
"links": links,
"metadata": metadata,
}
)
return self.steps
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Play the WikiRun game")
# Add mutual exclusion group for player type
player_group = parser.add_mutually_exclusive_group(required=True)
player_group.add_argument("--human", action="store_true", help="Play as a human")
player_group.add_argument("--agent", action="store_true", help="Use an AI agent to play")
# Game parameters
parser.add_argument("--start", type=str, default="British Library", help="Starting article title")
parser.add_argument("--end", type=str, default="Saint Lucia", help="Target article title")
parser.add_argument("--db", type=str, required=True, help="Path to SQLite database")
parser.add_argument("--max-steps", type=int, default=10, help="Maximum number of steps allowed (default: 10)")
# Agent parameters (only used with --agent)
parser.add_argument("--model", type=str, default="gpt-4o", help="Model to use for the agent (default: gpt-4o)")
parser.add_argument("--api-base", type=str, default="https://api.openai.com/v1",
help="API base URL (default: https://api.openai.com/v1)")
parser.add_argument("--max-links", type=int, default=200, help="Maximum number of links to consider (default: 200)")
parser.add_argument("--max-tries", type=int, default=3, help="Maximum number of tries for the agent (default: 3)")
parser.add_argument("--seed", type=int, default=None, help="Random seed for reproducibility")
args = parser.parse_args()
# Initialize the database
db = SQLiteDB(args.db)
# Initialize the player based on the argument
if args.human:
player = Player("Human")
else: # args.agent is True
player = AgentPlayer(
model=args.model,
api_base=args.api_base,
verbose=True,
max_links=args.max_links,
max_tries=args.max_tries,
target_article=args.end,
seed=args.seed
)
# Create and run the game
game = Game(
start_article=args.start,
target_article=args.end,
db=db,
max_allowed_steps=args.max_steps,
player=player,
verbose=True
)
steps = asyncio.run(game.run())
print(f"Game over in {len(steps)} steps")
for i, step in enumerate(steps):
print(f"Step {i}: {step['type']}")
print(f" Article: {step['article']}")
print(f" Links: {step.get('links', [])}")
print(f" Metadata: {step.get('metadata', {})}")
print("\n\n")