Update app.py

app.py

@@ -1,17 +1,19 @@
 import os
 import gradio as gr
 import google.generativeai as genai
+from google.generativeai.types import HarmCategory, HarmBlockThreshold
 from markdown_pdf import MarkdownPdf, Section
 
 # -------------------- CONFIG --------------------
-genai.configure(api_key=os.getenv("GEMINI_API_KEY"))
-
-# ---------- PROMPTS ----------
-
-# Student transcription prompt
-TRANSCRIPTION_PROMPT = """Your Role: You are an expert technical transcriber specializing in mathematical and scientific documents. 
+API_KEY = os.getenv("GOOGLE_AI_STUDIO_API_KEY")
+genai.configure(api_key=API_KEY)
+
+# ---------- PROMPTS IN JSON ----------
+PROMPTS = {
+    "TRANSCRIPTION_PROMPT": {
+        "role": "system",
+        "content": """Your Role: You are an expert technical transcriber specializing in mathematical and scientific documents. 
 Your mission is to convert handwritten solutions from a provided image or PDF into a clean, accurate, and logically structured Markdown format.
-
 Instructions:
 - Use ## for questions, ### for subquestions.
 - Transcribe only the corrected, final version of the solution (ignore scribbles, cancellations, mistakes).
@@ -19,11 +21,12 @@ Instructions:
 - If something is illegible, use [illegible].
 - Do not recreate graphs, only describe them.
 """
+    },
 
-# Markscheme transcription prompt
-MARKSCHEME_TRANSCRIPTION_PROMPT = """Your Role: You are an expert transcriber. 
+    "MARKSCHEME_TRANSCRIPTION_PROMPT": {
+        "role": "system",
+        "content": """Your Role: You are an expert transcriber. 
 Convert the official marking scheme from the provided PDF into clean, structured Markdown.
-
 Instructions:
 - Preserve all structure (questions, subquestions).
 - Keep M, A, R annotations exactly as written.
@@ -32,71 +35,61 @@ Instructions:
 - Format in Markdown using ## and ### for hierarchy.
 - Use code blocks for equations.
 """
+    },
 
-# Grading prompt with rules + red highlighting
-GRADING_PROMPT = """You are an official examiner. Use the following grading rules strictly.
-
+    "GRADING_PROMPT": {
+        "role": "system",
+        "content": """You are an official examiner. Use the following grading rules strictly.
 Abbreviations:  
 - M: Marks awarded for attempting to use a correct Method.  
 - A: Marks awarded for an Answer or for Accuracy; often dependent on preceding M marks.  
 - R: Marks awarded for clear Reasoning.  
 - AG: Answer given in the question and so no marks are awarded.  
 - FT: Follow through. The practice of awarding marks, despite candidate errors in previous parts, for their correct methods/answers using incorrect results.  
-
 --------------------------------------------  
 ## 1. General  
 Award marks using the annotations as noted in the markscheme (e.g., M1, A2).  
-
 ## 2. Method and Answer/Accuracy marks  
 - Do not automatically award full marks for a correct answer; all working must be checked.  
 - It is generally not possible to award M0 followed by A1.  
 - Where M and A marks are noted on the same line (M1A1), M is for method, A is for accuracy.  
 - Multiple A marks can be independent.  
-
 ## 3. Implied marks  
 Implied marks (M1) can only be awarded if correct work is seen or implied.  
-
 ## 4. Follow through (FT) marks  
 - Award FT if an earlier wrong answer is used consistently later.  
 - Do not award FT if the result contradicts the question (e.g., probability > 1).  
-
 ## 5. Mis-read (MR)  
 - Penalize once if the candidate misreads a value.  
 - Award other marks as appropriate.  
-
 ## 6. Alternative methods  
-- Accept valid alternatives unless "Hence" forbids it.  
-
+- Accept valid alternatives unless \"Hence\" forbids it.  
 ## 7. Alternative forms  
 - Accept equivalent numeric/algebraic forms unless specified otherwise.  
-
 ## 8. Format and accuracy of answers  
 - Use correct accuracy (3 s.f. if not specified).  
 - Arithmetic and algebra should be simplified.  
-
 ## 9. Presentation of candidate work  
 - Ignore crossed-out work unless indicated.  
 - Mark only the first solution unless candidate specifies otherwise.  
-
 --------------------------------------------  
-
 ### OUTPUT FORMAT
 Produce a GitHub-flavored Markdown table with 3 columns:  
-
 | Student wrote | Marks Awarded | Reason |  
 |---------------|---------------|--------|  
-
 Special Formatting Rule:  
-- Whenever a mark is lost (M0, A0, R0 etc.), wrap it in red using: `<span style="color:red">M0</span>`.  
+- Whenever a mark is lost (M0, A0, R0 etc.), wrap it in red using: `<span style=\"color:red\">M0</span>`.  
+- Also make the corresponding **Reason column text red** when a mark is lost.  
 - Keep awarded marks (M1, A1, etc.) in plain text.  
 - If mixed (e.g., M1A0A1), only highlight the lost marks (`A0`).  
-
 After the table, provide:  
 ### Summary & Final Mark  
 - Total marks obtained vs total available  
 - Any FT (follow-through) applied  
 - Classification of errors (Conceptual, Silly mistake, Misread, etc.)  
 """
+    }
+}
 
 # ---------- HELPER: Save to PDF ----------
 def save_as_pdf(text, filename="output.pdf"):
@@ -109,9 +102,16 @@ def save_as_pdf(text, filename="output.pdf"):
 def transcribe_student(ans_file):
     try:
         ans_uploaded = genai.upload_file(path=ans_file, display_name="Answer Sheet")
-        model = genai.GenerativeModel("gemini-2.5-pro", generation_config={"temperature": 0})
+        model = genai.GenerativeModel("gemini-2.5-pro-exp-03-25")
+
+        resp = model.generate_content(
+            [PROMPTS["TRANSCRIPTION_PROMPT"]["content"], ans_uploaded],
+            safety_settings={
+                HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
+                HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
+            }
+        )
 
-        resp = model.generate_content([TRANSCRIPTION_PROMPT, ans_uploaded])
         transcription = getattr(resp, "text", None)
         if not transcription and resp.candidates:
             transcription = resp.candidates[0].content.parts[0].text
@@ -125,9 +125,16 @@ def transcribe_student(ans_file):
 def transcribe_ms(ms_file):
     try:
         ms_uploaded = genai.upload_file(path=ms_file, display_name="Markscheme")
-        model = genai.GenerativeModel("gemini-2.5-pro", generation_config={"temperature": 0})
+        model = genai.GenerativeModel("gemini-2.5-pro-exp-03-25")
+
+        resp = model.generate_content(
+            [PROMPTS["MARKSCHEME_TRANSCRIPTION_PROMPT"]["content"], ms_uploaded],
+            safety_settings={
+                HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
+                HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
+            }
+        )
 
-        resp = model.generate_content([MARKSCHEME_TRANSCRIPTION_PROMPT, ms_uploaded])
         ms_transcription = getattr(resp, "text", None)
         if not ms_transcription and resp.candidates:
             ms_transcription = resp.candidates[0].content.parts[0].text
@@ -141,14 +148,17 @@ def transcribe_ms(ms_file):
 def grade(qp_file, ms_transcription, student_transcription):
     try:
         qp_uploaded = genai.upload_file(path=qp_file, display_name="Question Paper")
-        model = genai.GenerativeModel("gemini-2.5-pro", generation_config={"temperature": 0})
-
-        response = model.generate_content([
-            GRADING_PROMPT,
-            qp_uploaded,
-            "### Markscheme Transcription:\n" + ms_transcription,
-            "### Student Transcription:\n" + student_transcription
-        ])
+        model = genai.GenerativeModel("gemini-2.5-pro-exp-03-25")
+
+        response = model.generate_content(
+            [PROMPTS["GRADING_PROMPT"]["content"], qp_uploaded,
+             "### Markscheme Transcription:\n" + ms_transcription,
+             "### Student Transcription:\n" + student_transcription],
+            safety_settings={
+                HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
+                HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
+            }
+        )
 
         grading = getattr(response, "text", None)
         if not grading and response.candidates: