Rajarshi255 commited on
Commit
c25c1df
Β·
verified Β·
1 Parent(s): 2f79398

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +156 -217
app.py CHANGED
@@ -1,52 +1,36 @@
1
- import os, json, textwrap, re
2
  import gradio as gr
3
  from openai import OpenAI
4
 
5
- # ==================== Config ====================
6
- # DeepSeek (primary) β€” OpenAI-compatible API
7
- DEEPSEEK_API_KEY = os.getenv("DEEPSEEK_API_KEY")
8
  DEEPSEEK_BASE_URL = os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1")
9
- DEEPSEEK_MODEL = os.getenv("DEEPSEEK_MODEL", "DeepSeek-V3.2-Exp") # non-thinking
10
 
11
- # OpenAI (fallback)
12
  OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
13
- OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-5")
14
-
15
- # Limits (basic hygiene for web)
16
- MAX_CHARS = 12000 # limit paste size to keep latency sane
17
-
18
- # ==================== Prompts ====================
19
- RUBRIC_USER_TEXT = """
20
- You are a teacher grading a student's summary. Follow these rules strictly.
21
-
22
- Step 1: Decide if you KNOW the story.
23
- - If you KNOW the work (book/movie), set "knows_story": true.
24
- - If you DO NOT know the work, set "knows_story": false.
25
-
26
- Step 2: Scoring rules.
27
- - If knows_story=true, grade all of these:
28
- 1) Comprehension – 1 to 5 stars (how accurately the student captured the story or message).
29
- 2) Expression – 1 to 5 stars (clarity, coherence, fluency; how understandable if narrated).
30
- 3) Reflection/Moral Takeaway – 1 to 5 stars (thoughtfulness, personal connection).
31
- 4) Grammar (lenient): Identify grammar/spelling/punctuation mistakes in the student's text. Ignore minor conversational-English slips.
32
- Return a list of issues where each item contains:
33
- - "error_excerpt": student's exact snippet with the mistake (max ~120 chars)
34
- - "correction": corrected version of that snippet
35
- - "explanation": short reason (e.g., "subject–verb agreement", "punctuation spacing")
36
- - "start_index": 0-based char index in the ORIGINAL text where the error begins
37
- - "end_index": exclusive end index
38
- - If knows_story=false, omit Comprehension (set it to null) but grade the rest with the same grammar rules.
39
-
40
- Constraints:
41
- - Keep "error_excerpt" and "correction" minimal and tightly focused (no rewriting entire paragraphs).
42
- - Indices MUST refer to the student's ORIGINAL text provided.
43
- - After listing issues, also return a "corrected_summary" that lightly fixes the student's full passage (minimal edits, same meaning, natural).
44
-
45
- IMPORTANT: Do NOT compute "Length" stars. The app computes them locally by word count bins.
46
- """
47
 
48
- JSON_SCHEMA_HINT = """
49
- Return ONLY valid JSON with this EXACT structure:
 
 
 
 
 
 
 
 
50
  {
51
  "knows_story": true/false,
52
  "scores": {
@@ -55,35 +39,23 @@ Return ONLY valid JSON with this EXACT structure:
55
  "Reflection/Moral Takeaway": 1-5
56
  },
57
  "grammar_issues": [
58
- {
59
- "error_excerpt": "<string>",
60
- "correction": "<string>",
61
- "explanation": "<string>",
62
- "start_index": 0,
63
- "end_index": 0
64
- }
65
  ],
66
- "corrected_summary": "<the student's passage corrected for grammar (minimal edits)>",
67
- "remarks": "<1–2 sentence feedback (plain text)>"
68
  }
69
- - If there are no grammar issues, return "grammar_issues": [].
70
- - No markdown. No extra keys. Numbers only for scores.
71
  """
72
 
73
- # ==================== LLM Bridge ====================
74
- def llm_chat(messages, prefer_deepseek=True):
75
- """
76
- Try DeepSeek first (non-thinking mode via OpenAI-compatible API); on error, fallback to OpenAI.
77
- Return (content, model_used_str).
78
- """
79
- last_err = None
80
- if prefer_deepseek and DEEPSEEK_API_KEY:
81
- try:
82
  ds = OpenAI(api_key=DEEPSEEK_API_KEY, base_url=DEEPSEEK_BASE_URL)
83
  resp = ds.chat.completions.create(model=DEEPSEEK_MODEL, messages=messages)
84
  return resp.choices[0].message.content, f"DeepSeek ({DEEPSEEK_MODEL})"
85
- except Exception as e:
86
- last_err = e # fall through to OpenAI
87
 
88
  if OPENAI_API_KEY:
89
  try:
@@ -91,193 +63,160 @@ def llm_chat(messages, prefer_deepseek=True):
91
  resp = oa.chat.completions.create(model=OPENAI_MODEL, messages=messages)
92
  return resp.choices[0].message.content, f"OpenAI ({OPENAI_MODEL})"
93
  except Exception as e:
94
- last_err = e
95
 
96
- if not DEEPSEEK_API_KEY and not OPENAI_API_KEY:
97
- raise RuntimeError("No API key found. Set DEEPSEEK_API_KEY and/or OPENAI_API_KEY as Secrets.")
98
- raise RuntimeError(f"Both DeepSeek and OpenAI calls failed. Last error: {last_err}")
99
 
100
- # ==================== Helpers ====================
101
  def salvage_json(s: str) -> dict:
102
  s = s.strip()
103
  try:
104
  return json.loads(s)
105
- except json.JSONDecodeError:
106
  m = re.search(r"\{.*\}\s*$", s, flags=re.S)
107
- if not m:
108
- raise ValueError("Model did not return valid JSON.")
109
- return json.loads(m.group(0))
110
-
111
- def stars_bar(value: float | int | None, max_stars: int = 5) -> str:
112
- if value is None or float(value) <= 0:
113
- return "β€”"
114
- n_int = max(0, min(max_stars, int(round(float(value)))))
115
- return "⭐" * n_int
116
-
117
- def count_words(text: str) -> int:
118
- return len(re.findall(r"\b\w[\w'-]*\b", text))
119
-
120
- def length_stars(word_count: int) -> int:
121
- if word_count < 100: return 0
122
- if word_count <= 150: return 1
123
- if word_count <= 200: return 3
124
- if word_count <= 250: return 4
125
  return 5
126
 
127
- def clamp_1_to_5(x): return max(1, min(5, int(x)))
 
128
 
129
- def grammar_stars_from_count(n: int) -> float:
130
- if n >= 50: return 0.0
131
- return max(0.0, 5.0 - 0.1 * n)
 
 
132
 
133
- def annotate_original_with_issues(text: str, issues: list[dict]) -> str:
134
- clean = []
135
- for i, it in enumerate(issues, 1):
136
- try:
137
- si, ei = int(it["start_index"]), int(it["end_index"])
138
- if 0 <= si < ei <= len(text): clean.append((si, ei, i))
139
- except Exception:
140
- continue
141
- clean.sort(key=lambda x: x[0])
142
- out, cursor, last_end = [], 0, -1
143
- for si, ei, idx in clean:
144
- if si < last_end: # skip overlaps
145
- continue
146
- out.append(text[cursor:si])
147
- out.append(f"[[#{idx}: {text[si:ei]}]]")
148
- cursor, last_end = ei, ei
149
- out.append(text[cursor:])
150
- return "".join(out)
151
-
152
- def apply_corrections_by_offsets(text: str, issues: list[dict]) -> str:
153
- spans = []
154
- for it in issues:
155
- try:
156
- si, ei = int(it["start_index"]), int(it["end_index"])
157
- cor = str(it.get("correction", ""))
158
- if 0 <= si < ei <= len(text): spans.append((si, ei, cor))
159
- except Exception:
160
- continue
161
- spans.sort(key=lambda x: x[0])
162
- out, cursor, last_end = [], 0, -1
163
- for si, ei, cor in spans:
164
- if si < last_end: # skip overlaps
165
- continue
166
- out.append(text[cursor:si]); out.append(cor)
167
- cursor, last_end = ei, ei
168
- out.append(text[cursor:])
169
- return "".join(out)
170
-
171
- # ==================== Core grading ====================
172
- def grade_summary(title: str, student_summary: str):
173
- if not title.strip():
174
- return "⚠️ Please enter a title.", None
175
- if not student_summary.strip():
176
- return "⚠️ Please paste the student's summary.", None
177
- if len(student_summary) > MAX_CHARS:
178
- return f"⚠️ Summary too long ({len(student_summary)} chars). Limit is {MAX_CHARS}.", None
179
 
180
- prompt = f"""
181
- TITLE: {title}
 
 
182
 
183
- STUDENT SUMMARY (ORIGINAL TEXT BELOW):
184
- {student_summary}
185
 
186
- {RUBRIC_USER_TEXT}
 
187
 
188
- {JSON_SCHEMA_HINT}
189
  """
190
- # Call LLM
191
- content, model_used = llm_chat([{"role": "user", "content": prompt}], prefer_deepseek=True)
192
  data = salvage_json(content)
193
 
194
- # Extract
195
- knows_story = bool(data.get("knows_story", False))
196
- scores = data.get("scores", {}) or {}
197
- comp = scores.get("Comprehension", None)
198
- expr = clamp_1_to_5(scores.get("Expression", 1))
199
- refl = clamp_1_to_5(scores.get("Reflection/Moral Takeaway", 1))
200
- issues = data.get("grammar_issues", []) or []
201
- corrected = (data.get("corrected_summary") or "").strip()
202
- remarks = (data.get("remarks") or "").strip()
203
-
204
- if not knows_story:
205
- comp = None
206
-
207
- wc = count_words(student_summary)
208
- len_s = length_stars(wc)
209
  grammar_count = len(issues)
210
- grammar_star = round(grammar_stars_from_count(grammar_count), 1)
211
 
212
- # Average over included categories
213
- vals = [float(expr), float(refl), float(len_s), float(grammar_star)]
214
- if knows_story and comp is not None:
215
- vals.append(float(comp))
216
- avg = round(sum(vals) / len(vals), 2) if vals else 0.0
217
 
218
- # Compose Markdown report
219
  md = []
220
- md.append(f"### πŸ“š Summary Grading Report")
221
  md.append(f"**Title:** {title}")
222
- md.append(f"**Word count:** {wc}")
223
  md.append(f"**Model used:** `{model_used}`")
224
  md.append("")
225
- md.append("#### Category-wise Scores")
226
- if knows_story and comp is not None:
227
- md.append(f"- **Comprehension:** {stars_bar(comp)} ({comp} stars)")
228
- else:
229
- md.append(f"- **Comprehension:** *(omitted)*")
230
- md.append(f"- **Expression:** {stars_bar(expr)} ({expr} stars)")
231
- md.append(f"- **Reflection/Moral Takeaway:** {stars_bar(refl)} ({refl} stars)")
232
- md.append(f"- **Length:** {stars_bar(len_s)} ({len_s} stars, based on {wc} words)")
233
- md.append(f"- **Grammar:** {stars_bar(grammar_star)} ({grammar_star:.1f} stars, {grammar_count} mistakes)")
234
- md.append("")
235
- md.append(f"**Average Score:** {avg:.2f} {stars_bar(avg)} *(out of 5)*")
236
- md.append("")
237
- md.append("#### Corrected Passage (LLM, minimal edits)")
238
- md.append(f"{corrected if corrected else '_(No changes returned by the model.)_'}")
239
  md.append("")
240
- md.append("#### Annotated Original (exact mistake spans)")
241
- annotated = annotate_original_with_issues(student_summary, issues)
242
- md.append(f"{annotated}")
243
  md.append("")
244
- md.append("#### Grammar Issues (with positions)")
245
- if not issues:
246
- md.append("- None πŸŽ‰")
247
- else:
248
- for i, it in enumerate(issues, 1):
249
- err = (it.get('error_excerpt') or '').strip()
250
- cor = (it.get('correction') or '').strip()
251
- why = (it.get('explanation') or '').strip()
252
- si = it.get('start_index', '?')
253
- ei = it.get('end_index', '?')
254
- md.append(f"{i}. `[{si}:{ei}]` **{err}** β†’ **{cor}** β€” {why}")
255
  md.append("")
256
- md.append("#### Program-applied corrections (by offsets)")
257
- program_fixed = apply_corrections_by_offsets(student_summary, issues)
258
- md.append(f"{program_fixed}")
259
- md.append("")
260
- md.append("#### Overall Remark")
261
- md.append(f"{remarks if remarks else '_(None)_'}")
262
- md_str = "\n".join(md)
 
263
 
264
- # Also return plain corrected text as a second output (easy copy)
265
- return md_str, corrected
266
 
267
  # ==================== UI ====================
268
- with gr.Blocks(title="Summary Grader (DeepSeek β†’ OpenAI fallback)") as demo:
269
- gr.Markdown("## Summary Grader\nDeepSeek-V3.2-Exp first; falls back to OpenAI GPT-5 if needed.")
 
270
  with gr.Row():
271
- title_in = gr.Textbox(label="Book/Movie Title", placeholder="The Gift of the Magi")
272
- summary_in = gr.Textbox(label="Paste the student's summary", lines=12, placeholder="Paste text here...")
273
- btn = gr.Button("Grade Summary", variant="primary")
 
 
 
274
  with gr.Row():
275
- report_out = gr.Markdown(label="Grading Report")
276
- corrected_out = gr.Textbox(label="Corrected Passage (plain text)", lines=10)
 
 
 
 
277
 
278
- btn.click(fn=grade_summary, inputs=[title_in, summary_in], outputs=[report_out, corrected_out])
 
 
279
 
280
- # Entrypoint for Spaces
281
  if __name__ == "__main__":
282
  demo.launch()
283
-
 
1
+ import os, json, textwrap, re, time
2
  import gradio as gr
3
  from openai import OpenAI
4
 
5
+ # ==================== LLM CONFIG ====================
6
+ DEEPSEEK_API_KEY = os.getenv("DEEPSEEK_API_KEY")
 
7
  DEEPSEEK_BASE_URL = os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com/v1")
8
+ DEEPSEEK_MODEL = os.getenv("DEEPSEEK_MODEL", "DeepSeek-V3.2-Exp")
9
 
 
10
  OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
11
+ OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-5")
12
+
13
+ # ==================== GOOGLE SHEETS CONFIG ====================
14
+ SVC_JSON = os.getenv("GOOGLE_SERVICE_ACCOUNT_JSON") # full JSON as string
15
+ SHEET_ID = os.getenv("SHEETS_SPREADSHEET_ID")
16
+ WS_ID = os.getenv("SHEETS_WORKSHEET_ID")
17
+
18
+ MAX_CHARS = 12000
19
+
20
+ # ==================== RUBRIC ====================
21
+ RUBRIC_PROMPT = """
22
+ You are a teacher grading a student's summary. Follow these steps:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
 
24
+ 1. Decide if you KNOW the story.
25
+ - If you know it, "knows_story": true. Otherwise false.
26
+
27
+ 2. Grade the following:
28
+ - Comprehension (1–5 stars) – only if knows_story=true
29
+ - Expression (1–5 stars)
30
+ - Reflection/Moral Takeaway (1–5 stars)
31
+ - Grammar (lenient): identify grammar/spelling/punctuation issues; ignore casual or stylistic ones.
32
+
33
+ Return valid JSON:
34
  {
35
  "knows_story": true/false,
36
  "scores": {
 
39
  "Reflection/Moral Takeaway": 1-5
40
  },
41
  "grammar_issues": [
42
+ {"error_excerpt": "...", "correction": "...", "explanation": "...", "start_index": 0, "end_index": 0}
 
 
 
 
 
 
43
  ],
44
+ "corrected_summary": "<corrected passage>",
45
+ "remarks": "<brief remark>"
46
  }
 
 
47
  """
48
 
49
+ # ==================== LLM BRIDGE ====================
50
+ def llm_chat(messages):
51
+ """Try DeepSeek first, then OpenAI."""
52
+ try:
53
+ if DEEPSEEK_API_KEY:
 
 
 
 
54
  ds = OpenAI(api_key=DEEPSEEK_API_KEY, base_url=DEEPSEEK_BASE_URL)
55
  resp = ds.chat.completions.create(model=DEEPSEEK_MODEL, messages=messages)
56
  return resp.choices[0].message.content, f"DeepSeek ({DEEPSEEK_MODEL})"
57
+ except Exception as e:
58
+ print("DeepSeek failed:", e)
59
 
60
  if OPENAI_API_KEY:
61
  try:
 
63
  resp = oa.chat.completions.create(model=OPENAI_MODEL, messages=messages)
64
  return resp.choices[0].message.content, f"OpenAI ({OPENAI_MODEL})"
65
  except Exception as e:
66
+ print("OpenAI failed:", e)
67
 
68
+ raise RuntimeError("No LLM API available. Please set DEEPSEEK_API_KEY or OPENAI_API_KEY.")
 
 
69
 
 
70
  def salvage_json(s: str) -> dict:
71
  s = s.strip()
72
  try:
73
  return json.loads(s)
74
+ except:
75
  m = re.search(r"\{.*\}\s*$", s, flags=re.S)
76
+ return json.loads(m.group(0)) if m else {"error": "bad_json"}
77
+
78
+ def stars_bar(val):
79
+ if not val or val <= 0: return "β€”"
80
+ return "⭐" * int(round(val))
81
+
82
+ def count_words(text): return len(re.findall(r"\b\w[\w'-]*\b", text))
83
+
84
+ def length_stars(wc):
85
+ if wc < 100: return 0
86
+ if wc <= 150: return 1
87
+ if wc <= 200: return 3
88
+ if wc <= 250: return 4
 
 
 
 
 
89
  return 5
90
 
91
+ def grammar_stars(n):
92
+ return 0 if n >= 50 else round(5 - 0.1 * n, 1)
93
 
94
+ # ==================== GOOGLE SHEETS SAVE ====================
95
+ def append_to_sheet(row: list):
96
+ """Try to append row. If not configured, skip gracefully."""
97
+ if not (SVC_JSON and SHEET_ID and WS_ID):
98
+ return "πŸ’Ύ Save skipped: Google Sheets not configured."
99
 
100
+ try:
101
+ import gspread
102
+ from google.oauth2.service_account import Credentials
103
+ creds = Credentials.from_service_account_info(json.loads(SVC_JSON),
104
+ scopes=["https://www.googleapis.com/auth/spreadsheets"])
105
+ gc = gspread.authorize(creds)
106
+ ws = gc.open_by_key(SHEET_ID).get_worksheet_by_id(int(WS_ID))
107
+ ws.append_row(row, value_input_option="USER_ENTERED")
108
+ return "βœ… Saved to Google Sheet."
109
+ except Exception as e:
110
+ return f"❌ Save failed: {e}"
111
+
112
+ # ==================== INTRO MESSAGE ====================
113
+ def intro_message(name, klass):
114
+ if not name.strip() or not klass.strip():
115
+ return "πŸ‘‹ Please tell me your name and class first!"
116
+ return f"""
117
+ ### 🌟 Hi {name} of {klass}!
118
+ Welcome to the **Summary Grader!**
119
+
120
+ Here’s how I’ll mark your story:
121
+
122
+ 1️⃣ **Comprehension** – Did you really understand the story?
123
+ 2️⃣ **Expression** – How clearly you write and explain.
124
+ 3️⃣ **Reflection / Moral** – What did you learn or feel?
125
+ 4️⃣ **Grammar & Neatness** – Spelling and punctuation.
126
+ 5️⃣ **Length** –
127
+ - Less than 100 words β†’ 0⭐
128
+ - 100–150 words β†’ 1⭐
129
+ - 150–200 words β†’ 3⭐
130
+ - 200–250 words β†’ 4⭐
131
+ - 250+ words β†’ 5⭐
132
+
133
+ 🌈 Write in your own words, think deeply, and use neat sentences.
134
+ Paste your story below when ready and click **Grade My Work!**
135
+ """
 
 
 
 
 
 
 
 
 
 
136
 
137
+ # ==================== MAIN GRADER ====================
138
+ def grade_summary(name, klass, title, summary):
139
+ if not all([name.strip(), klass.strip(), title.strip(), summary.strip()]):
140
+ return "⚠️ Please fill all fields.", "", ""
141
 
142
+ prompt = f"""
143
+ Title: {title}
144
 
145
+ Student Summary:
146
+ {summary}
147
 
148
+ {RUBRIC_PROMPT}
149
  """
150
+ content, model_used = llm_chat([{"role": "user", "content": prompt}])
 
151
  data = salvage_json(content)
152
 
153
+ knows_story = data.get("knows_story", False)
154
+ scores = data.get("scores", {})
155
+ comp = scores.get("Comprehension")
156
+ expr = scores.get("Expression", 1)
157
+ refl = scores.get("Reflection/Moral Takeaway", 1)
158
+ issues = data.get("grammar_issues", [])
159
+ corrected = data.get("corrected_summary", "")
160
+ remarks = data.get("remarks", "")
161
+
162
+ wc = count_words(summary)
163
+ len_star = length_stars(wc)
 
 
 
 
164
  grammar_count = len(issues)
165
+ grammar_star = grammar_stars(grammar_count)
166
 
167
+ values = [expr, refl, len_star, grammar_star]
168
+ if knows_story and comp: values.append(comp)
169
+ avg = round(sum(values) / len(values), 2)
 
 
170
 
 
171
  md = []
172
+ md.append(f"### πŸ“š Summary Report for {name} (Class {klass})")
173
  md.append(f"**Title:** {title}")
 
174
  md.append(f"**Model used:** `{model_used}`")
175
  md.append("")
176
+ md.append("#### ⭐ Category-wise Scores")
177
+ md.append(f"- Comprehension: {stars_bar(comp)} ({comp if comp else 'β€”'})")
178
+ md.append(f"- Expression: {stars_bar(expr)} ({expr})")
179
+ md.append(f"- Reflection/Moral: {stars_bar(refl)} ({refl})")
180
+ md.append(f"- Length: {stars_bar(len_star)} ({len_star} from {wc} words)")
181
+ md.append(f"- Grammar: {stars_bar(grammar_star)} ({grammar_star} from {grammar_count} mistakes)")
 
 
 
 
 
 
 
 
182
  md.append("")
183
+ md.append(f"**Final Average:** {avg} {stars_bar(avg)} out of 5 ⭐")
 
 
184
  md.append("")
185
+ md.append(f"**Remarks:** {remarks}")
 
 
 
 
 
 
 
 
 
 
186
  md.append("")
187
+ md.append(f"**Corrected passage:**\n\n{corrected}")
188
+ report = "\n".join(md)
189
+
190
+ # Google Sheet save
191
+ row = [time.strftime("%Y-%m-%d %H:%M:%S"), name, klass, title, wc,
192
+ "Yes" if knows_story else "No", comp or "", expr, refl,
193
+ len_star, grammar_count, grammar_star, avg, model_used, remarks]
194
+ status = append_to_sheet(row)
195
 
196
+ return report, corrected, status
 
197
 
198
  # ==================== UI ====================
199
+ with gr.Blocks(title="Summary Grader (DeepSeek β†’ OpenAI)") as demo:
200
+ gr.Markdown("## ✨ Summary Grader\nDeepSeek first, OpenAI fallback. Saves to Google Sheet if configured.")
201
+
202
  with gr.Row():
203
+ name = gr.Textbox(label="Student Name")
204
+ klass = gr.Textbox(label="Class/Grade")
205
+
206
+ intro_out = gr.Markdown()
207
+ gr.Button("Show How I’ll Mark You").click(intro_message, [name, klass], intro_out)
208
+
209
  with gr.Row():
210
+ title = gr.Textbox(label="Book/Movie Title", placeholder="The Gift of the Magi")
211
+ summary = gr.Textbox(label="Your Summary", lines=10, placeholder="Paste here...")
212
+
213
+ report_out = gr.Markdown()
214
+ corrected_out = gr.Textbox(label="Corrected Text")
215
+ status_out = gr.Markdown()
216
 
217
+ gr.Button("Grade My Work!").click(grade_summary,
218
+ [name, klass, title, summary],
219
+ [report_out, corrected_out, status_out])
220
 
 
221
  if __name__ == "__main__":
222
  demo.launch()