README modified with clarifications

README.md

@@ -18,29 +18,32 @@ short_description: Final Assignment for agents course
 
 The graph is built with LangGraph framework. There is a control of iterations between the agent and the tools node, with a field in the state. The **route_tools** function checks if the **iterations** number, updated in the **increase** node, is greater than the stablished in MAX_ITERATIONS constant.
 
-```mermaid
----
-config:
-  flowchart:
-    curve: linear
----
-graph TD;
-	__start__([<p>__start__</p>]):::first
-	input(input)
-	agent(agent)
-	increase(increase)
-	tools(tools)
-	final_output(final_output)
-	__end__([<p>__end__</p>]):::last
-	__start__ --> input;
-	agent -. &nbsp;final_output&nbsp; .-> final_output;
-	agent -. &nbsp;tools&nbsp; .-> increase;
-	increase --> tools;
-	input --> agent;
-	tools --> agent;
-	final_output --> __end__;
-	classDef default fill:#f2f0ff,line-height:1.2
-	classDef first fill-opacity:0
-	classDef last fill:#bfb6fc
-
-```
+### Models
+I have tried two different models so far.
+* **gemini-2.5-flash**. This is free to use, taking advantage of the generous limits provided by Google AI for developers.
+* **gpt-oss-120b**. I have used it through HuggingFace inference providers. Since I have a pro account, 2$/month is more than enough to develop the Final Assignment project.
+This model, **gpt-oss-120b, does not work through _Together_ inference provider**. It seems that Together is not performing well with LangGraph. It worked just fine through **_Fireworks_** inference provider.
+
+### Tools
+* python_tool
+* reverse_tool
+* excel_file_to_markdown
+* sum_numbers
+* web_search
+* get_wikipedia_info
+* ask_audio_model
+* chess_tool (this one cannot run on HuggingFace Space)
+
+
+### Results
+
+The answers of the two models were cached after generation, and then submitted. I modified the gradio app to do so, as suggested in the template comments
+
+* gemini-2.5-flash. 40%
+* gpt-oss-120b. 60%
+* combined (taking correct results from both models): 65%. Only one additional answer provided by gemini.
+
+
+
+
+

agent.py

@@ -17,8 +17,8 @@ from tools import (
     )
 from chess_tool import chess_tool
 
-MODEL_PROVIDER = "gemini" 
-# MODEL_PROVIDER = "openai" 
+# MODEL_PROVIDER = "gemini"
+MODEL_PROVIDER = "openai" 
 
 MAX_ITERATIONS = 5
 
@@ -38,13 +38,14 @@ llm_gemini = ChatGoogleGenerativeAI(
     model="gemini-2.5-flash",
     include_thoughts=False,
     temperature=0,
-    max_output_tokens=256,
+    max_output_tokens=None,
     timeout=60,  # The maximum number of seconds to wait for a response.
     max_retries=2,
 )
 
 llm_openai = ChatOpenAI(
-    model="openai/gpt-oss-120b:together",
+    # model="openai/gpt-oss-120b:together",
+    model="openai/gpt-oss-120b:fireworks-ai",
     temperature=0,
     max_tokens=None, # type: ignore
     timeout=60,

chess_tool.py

@@ -15,7 +15,7 @@ def chess_tool(task_id: str, color_to_move: str) -> str:
     Given an image of a chessboard, and the color to move, 
     predict the FEN notation and suggest the best move.
     https://en.wikipedia.org/wiki/Forsyth%E2%80%93Edwards_Notation
-    
+    Important: the output of this tool is not to be modified in any way.
     Args:
         task_id (str): The identifier for the chessboard image.
         color_to_move (str): 'w' or 'b', the color to move ('white' or 'black').
@@ -44,5 +44,7 @@ def chess_tool(task_id: str, color_to_move: str) -> str:
     stockfish.set_fen_position(fen)
     best_move = stockfish.get_best_move()
     logger.info(f"Best move determined: {best_move!r}")
+    piece = stockfish.get_what_is_on_square(best_move[:2])  # type:ignore
+    next_move_fen = piece.value.upper() + best_move[2:] # type:ignore
 
-    return best_move #type: ignore
+    return next_move_fen #type: ignore