carolinacon's picture
updated chess tool and prompting
a4b0424
import base64
import json
import os
from typing import Dict, List, Optional, Any
import chess
import chess.engine
from langchain.chat_models import init_chat_model
from langchain.schema import SystemMessage, HumanMessage
from langchain.tools import tool
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
from utils.prompt_manager import prompt_mgmt
def encode_image_to_base64(image_path: str) -> str:
"""Encode image to base64 for API consumption"""
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode('utf-8')
class ChessPiecePosition(BaseModel):
"""Model for chess piece position"""
square: str = Field(..., description="Chess square notation (e.g., 'e4', 'a1')")
piece: str = Field(..., description="Piece type and color (e.g., 'white_king', 'black_queen')")
class ChessBoardAnalysis(BaseModel):
"""Model for complete chess board analysis"""
positions: List[ChessPiecePosition] = Field(..., description="List of all piece positions on the board")
def add_positions(self, positions: List[ChessPiecePosition]) -> None:
"""Add multiple positions to the analysis"""
for position in positions:
self.positions.append(position)
def merge_with(self, other: 'ChessBoardAnalysis') -> None:
"""Merge another analysis into this one (overwriting conflicts)"""
self.add_positions(other.positions)
def to_fen(self, active_color) -> str:
"""Convert the analysis to FEN notation (simplified)"""
# Create an 8x8 board representation
board = [['' for _ in range(8)] for _ in range(8)]
for position in self.positions:
file_idx = ord(position.square[0]) - ord('a')
rank_idx = 8 - int(position.square[1])
if 0 <= file_idx < 8 and 0 <= rank_idx < 8:
piece_char = self._piece_to_char(position.piece)
board[rank_idx][file_idx] = piece_char
# Convert to FEN string
fen_rows = []
for row in board:
fen_row = ''
empty_count = 0
for cell in row:
if cell == '':
empty_count += 1
else:
if empty_count > 0:
fen_row += str(empty_count)
empty_count = 0
fen_row += cell
if empty_count > 0:
fen_row += str(empty_count)
fen_rows.append(fen_row)
piece_placement = '/'.join(fen_rows)
# Determine active color
active_color_char = 'w' if active_color.lower() == 'white' else 'b'
# Build complete FEN string
castling_rights = "-"
en_passant = "-"
halfmove_clock = 0
fullmove_number = 1
fen_parts = [
piece_placement,
active_color_char,
castling_rights,
en_passant,
str(halfmove_clock),
str(fullmove_number)
]
return ' '.join(fen_parts)
def _piece_to_char(self, piece: str) -> str:
"""Convert piece description to FEN character"""
color, piece_type = piece.split('_')
piece_chars = {
'king': 'K', 'queen': 'Q', 'rook': 'R',
'bishop': 'B', 'knight': 'N', 'pawn': 'P'
}
char = piece_chars.get(piece_type, '')
return char.lower() if color == 'black' else char
class ChessVisionAnalyzer:
def __init__(self):
self.llm1 = init_chat_model(model="openai:gpt-4.1", temperature=0.0)
self.llm2 = ChatGoogleGenerativeAI(model="gemini-2.5-flash")
def analyze_board_orientation(self, active_color: str, image_path: str) -> str:
"""Analyze chess board image and return FEN notation"""
base64_image = encode_image_to_base64(image_path)
messages = [
SystemMessage(
content=prompt_mgmt.render_template("chess_board_orientation", {})),
HumanMessage(content=[
{
"type": "text",
"text": f"Analyze this chess board image and return the chess board orientation. I know that the "
f"active color is {active_color}"
},
{
"type": "image_url",
"image_url": {
"url": f"data:image/jpeg;base64,{base64_image}", "detail": "high"
}
}
])
]
response = self.llm1.invoke(messages)
return response.content
def analyze_board_from_image(self, board_orientation: str, image_path: str, llm_no: int,
squares: Optional[list] = None) -> Optional[ChessBoardAnalysis]:
"""Analyze chess board image and return FEN notation"""
base64_image = encode_image_to_base64(image_path)
squares_text = ""
if squares:
squares_text = (f"Focus only on these pieces {sorted(squares)} "
f"*** Important: make sure you detect correctly their position as they were challanged by another model. Take into account the board "
f"orientation."
)
messages = [
SystemMessage(
content=prompt_mgmt.render_template("chess_board_detection", {})),
HumanMessage(content=[
{
"type": "text",
"text": f"""Analyze this chess board image and return the pieces positions.
{board_orientation}
{squares_text}
Return the positions of the pieces in JSON format.
Use the following schema for each piece:
[{{
"square": "chess notation (e.g., 'e4', 'a1')",
"piece": "color_piece (e.g., 'white_king', 'black_queen')"
}},...
{{
"square": "chess notation (e.g., 'e4', 'a1')",
"piece": "color_piece (e.g., 'white_king', 'black_queen')"
}}
]
Very Important: Return only this list!
"""
},
{
"type": "image_url",
"image_url": {
"url": f"data:image/jpeg;base64,{base64_image}", "detail": "high"
}
}
])
]
if llm_no == 1:
response = self.llm1.invoke(messages)
else:
response = self.llm2.invoke(messages)
return self._parse_llm_response(response.content)
def analyze_board(self, active_color: str, file_reference: str) -> str:
board_orientation = self.analyze_board_orientation(active_color, file_reference)
first_analysis_res = self.analyze_board_from_image(board_orientation, file_reference, 1)
second_analysis_res = self.analyze_board_from_image(board_orientation, file_reference, 2)
result = self.compare_analyses(first_analysis_res, second_analysis_res)
if result['conflicts'] is not None and len(result['conflicts']) > 0:
arbitrage_result = self.arbitrate_conflicts(result, board_orientation, file_reference, 3)
return arbitrage_result.get("consensus").to_fen(active_color)
else:
result.get("consensus").to_fen(active_color)
def _parse_llm_response(self, response: str) -> Optional[ChessBoardAnalysis]:
"""Parse LLM response into ChessBoardAnalysis"""
try:
# Extract JSON from response
json_str = response.strip()
if "```json" in json_str:
json_str = json_str.split("```json")[1].split("```")[0].strip()
elif "```" in json_str:
json_str = json_str.split("```")[1].split("```")[0].strip()
data = json.loads(json_str)
print(data)
# Filter out items with null or None positions
positions = []
for item in data:
if item["piece"]:
positions.append(ChessPiecePosition(**item))
return ChessBoardAnalysis(positions=positions)
except Exception as e:
print(f"Failed to parse LLM response: {e}")
return None
def compare_analyses(self, analysis_1: ChessBoardAnalysis, analysis_2: ChessBoardAnalysis) -> dict:
"""Compare the given analyses and identify conflicts"""
if not analysis_1 or not analysis_2:
return {"conflicts": [], "consensus": None, "need_arbitration": False}
# Convert to dictionaries for easier comparison
dict_1 = {pos.square: pos.piece for pos in analysis_1.positions}
dict_2 = {pos.square: pos.piece for pos in analysis_2.positions}
conflicts = []
consensus = []
# Check all squares
all_squares = set(dict_1.keys()) | set(dict_2.keys())
for square in all_squares:
piece_1 = dict_1.get(square)
piece_2 = dict_2.get(square)
if piece_1 == piece_2:
if piece_1: # Only add if there's actually a piece
consensus.append(ChessPiecePosition(square=square, piece=piece_1))
else:
conflicts.append({
"square": square,
"analysis_1": piece_1,
"analysis_2": piece_2
})
need_arbitration = len(conflicts) > 0
return {
"conflicts": conflicts,
"consensus": ChessBoardAnalysis(positions=consensus),
"need_arbitration": need_arbitration
}
def arbitrate_conflicts(self, state: dict, board_orientation: str, image_path: str, depth: int = 1) -> dict:
"""Arbitrate conflicting piece positions"""
print(f"Arbitrating conflicts with depth {depth}")
conflicts = state.get("conflicts", [])
conflicts_sqares = []
for conflict in conflicts:
if conflict["analysis_1"] is not None: conflicts_sqares.append(conflict["analysis_1"])
if conflict["analysis_2"] is not None: conflicts_sqares.append(conflict["analysis_2"])
conflicts_sqares = set(conflicts_sqares)
print("Pieces with conflicts:", conflicts_sqares)
first_analysis_res = self.analyze_board_from_image(board_orientation, image_path, 1, conflicts_sqares)
second_analysis_res = self.analyze_board_from_image(board_orientation, image_path, 2, conflicts_sqares)
result = self.compare_analyses(first_analysis_res, second_analysis_res)
result.get("consensus").merge_with(state.get("consensus"))
if result['conflicts'] is not None and len(result['conflicts']) > 0:
if depth > 0:
depth -= 1
result = self.arbitrate_conflicts(result, board_orientation, image_path, depth)
else:
print("Arbitrage completed with conflicts. took llm2 as ground truth")
result.get("consensus").merge_with(second_analysis_res)
return result
class ChessEngineAnalyzer:
def __init__(self, stockfish_path: str = "stockfish"):
self.engine = chess.engine.SimpleEngine.popen_uci(stockfish_path)
def analyze_position(self, fen: str, depth: int = 18) -> Dict[str, Any]:
"""Analyze chess position using Stockfish"""
board = chess.Board(fen)
# Get top moves analysis
info = self.engine.analyse(board, chess.engine.Limit(depth=depth))
best_move = info.get("pv", [])[0] if info.get("pv") else None
evaluation = info.get("score", chess.engine.PovScore(chess.engine.Cp(0), chess.WHITE))
return {
"best_move": best_move.uci() if best_move else None,
"evaluation": str(evaluation),
"depth": depth,
"analysis": info
}
def close(self):
self.engine.quit()
class ChessMoveExplainer:
def __init__(self):
self.llm = ChatOpenAI(
model="gpt-4"
)
def explain_move(self, fen: str, move: str, analysis: Dict) -> str:
"""Generate human-readable explanation of the recommended move"""
board = chess.Board(fen)
san_move = board.san(chess.Move.from_uci(move))
prompt = f"""
Chess position FEN: {fen}
Recommended move: {san_move} ({move})
Engine evaluation: {analysis['evaluation']}
Analysis depth: {analysis['depth']}
Explain this move recommendation in simple terms. Consider:
1. Why this move is strong
2. What threats it creates or prevents
3. The strategic implications
4. Alternative moves and why they're inferior
5. Keep it concise but informative for an intermediate player
"""
response = self.llm.invoke([HumanMessage(content=prompt)])
return response.content
@tool
def chess_analysis_tool(active_color: str, file_reference: str) -> str:
"""
Tool for analyzing a chess board images and recommending moves
:param active_color: The color that should execute the next move
:param file_reference: the reference of the image to be analyzed
:return: the recommended move along with an analysis
"""
vision_analyzer = ChessVisionAnalyzer()
engine_analyzer = ChessEngineAnalyzer(os.getenv("CHESS_ENGINE_PATH"))
move_explainer = ChessMoveExplainer()
fen = vision_analyzer.analyze_board(active_color, file_reference)
print(f"Got fen {fen}")
analysis_result = engine_analyzer.analyze_position(fen)
print(f"Got analysis reslut {analysis_result}")
engine_analyzer.close()
explanation = move_explainer.explain_move(fen, analysis_result["best_move"], analysis_result)
return explanation