benjamin5607 commited on
Commit
9d0be3c
·
verified ·
1 Parent(s): f3b0f37

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +167 -113
app.py CHANGED
@@ -1,7 +1,7 @@
1
  import streamlit as st
2
  import feedparser
3
  import urllib.parse
4
- from groq import Groq
5
  import pandas as pd
6
  import json
7
  import os
@@ -23,7 +23,15 @@ if not api_key:
23
  st.stop()
24
 
25
  client = Groq(api_key=api_key)
26
- CURRENT_MODEL = "llama-3.3-70b-versatile"
 
 
 
 
 
 
 
 
27
 
28
  # --- 3. Configuration ---
29
  REGION_MAP = {
@@ -34,6 +42,31 @@ REGION_MAP = {
34
  }
35
 
36
  # --- 4. Core Functions ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
  def fetch_extensive_news(query, geo="US", limit=60, period="7d"):
38
  time_filter = f" when:{period}"
39
  encoded = urllib.parse.quote(query + time_filter)
@@ -59,7 +92,7 @@ def render_gauge(score, title):
59
  fig.update_layout(height=180, margin=dict(l=10, r=10, t=40, b=10), paper_bgcolor="rgba(0,0,0,0)")
60
  return fig
61
 
62
- # --- 5. 🚨 GLOBAL WAR ROOM (Includes Event Summary) ---
63
  st.title("🚨 OMNI-WATCH: T&S WAR ROOM")
64
  st.caption(f"Policy Risk Monitoring: {datetime.now().strftime('%H:%M:%S')} UTC")
65
 
@@ -69,23 +102,31 @@ if st.button("🔄 Refresh"):
69
  if 'hot_issues' not in st.session_state:
70
  with st.spinner("Scanning for Community Standard Violations..."):
71
  raw_news = fetch_extensive_news("violence OR hate speech OR disinformation OR scandal OR protest", limit=40, period="12h")
72
- news_context = "\n".join([f"Event: {n['title']}" for n in raw_news])
73
-
74
- # 🚨 PROMPT UPDATE: 첫 줄에 반드시 'Event Summary' 포함
75
- system_msg = """
76
- Identify TOP 3 incidents with highest 'Trust & Safety' risk.
77
- The 'summary' array MUST follow this order:
78
- 1. "Event: [Brief but clear summary of what happened]"
79
- 2. "Violation: [Specific Community Guideline breached]"
80
- 3. "Risk: [Potential offline harm]"
81
 
82
- Return ONLY JSON:
83
- {"issues": [{"title": "Short Title", "score": 85, "summary": ["Event: ...", "Violation: ...", "Risk: ..."], "link": ".."}]}
84
- """
85
- try:
86
- res = client.chat.completions.create(model=CURRENT_MODEL, messages=[{"role": "system", "content": system_msg}, {"role": "user", "content": news_context}], response_format={"type": "json_object"})
87
- st.session_state.hot_issues = json.loads(res.choices[0].message.content).get('issues', [])[:3]
88
- except: st.session_state.hot_issues = []
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
 
90
  if st.session_state.hot_issues:
91
  cols = st.columns(3)
@@ -114,56 +155,62 @@ with tab1:
114
  for geo in REGION_MAP[reg][:2]:
115
  all_news.extend(fetch_extensive_news(keyword, geo=geo, limit=6, period="7d"))
116
 
117
- news_summary = "\n".join([n['title'] for n in all_news[:45]])
118
-
119
- # 🚨 PROMPT UPDATE: 서머리에 '사건 개요' 필수 포함
120
- global_prompt = f"""
121
- Analyze '{keyword}' focusing on 'Community Guidelines' and 'Social Impact'.
122
-
123
- 1. Executive Summary: Start with a clear **Event Summary** of what happened. Then, analyze the systemic policy risks and societal harm. (300+ words).
124
- 2. Risk Landscape: Map findings to specific violations.
125
-
126
- Return ONLY JSON:
127
- {{
128
- "executive_summary": "Start with [The Event], then move to [Policy Analysis].",
129
- "risk_landscape": [
130
- {{"Component": "Primary Violation", "Findings": "...", "Risk_Level": "High"}},
131
- {{"Component": "Vulnerable Target", "Findings": "...", "Risk_Level": "Critical"}},
132
- {{"Component": "Offline Harm", "Findings": "...", "Risk_Level": "High"}},
133
- {{"Component": "Enforcement Gap", "Findings": "...", "Risk_Level": "Medium"}}
134
- ],
135
- "platform_intelligence": [
136
- {{"Platform": "TikTok", "Assessment": "...", "Strategy": "..."}},
137
- {{"Platform": "YouTube", "Assessment": "...", "Strategy": "..."}},
138
- {{"Platform": "Meta", "Assessment": "...", "Strategy": "..."}},
139
- {{"Platform": "X", "Assessment": "...", "Strategy": "..."}}
140
- ],
141
- "strategic_conclusion": "Final Trust & Safety recommendation."
142
- }}
143
- """
144
- res = client.chat.completions.create(model=CURRENT_MODEL, messages=[{"role": "system", "content": global_prompt}, {"role": "user", "content": news_summary}], response_format={"type": "json_object"})
145
- g_data = json.loads(res.choices[0].message.content)
146
-
147
- with st.container(border=True):
148
- st.subheader("1. Policy Impact Executive Summary")
149
- st.markdown(g_data.get('executive_summary').replace(". ", ".\n\n"))
150
-
151
- st.subheader("2. Guideline Violation Matrix")
152
- st.dataframe(pd.DataFrame(g_data.get('risk_landscape')), hide_index=True, use_container_width=True)
153
-
154
- st.subheader("3. Platform Enforcement Strategy")
155
- st.table(pd.DataFrame(g_data.get('platform_intelligence')))
156
-
157
- st.success(f"**T&S Recommendation:** {g_data.get('strategic_conclusion')}")
158
-
159
- # DOWNLOAD
160
- report_md = f"# OMNI-WATCH POLICY REPORT: {keyword.upper()}\n\n"
161
- report_md += f"## 1. EXECUTIVE SUMMARY\n{g_data.get('executive_summary')}\n\n"
162
- report_md += "## 2. VIOLATION LANDSCAPE\n"
163
- for item in g_data.get('risk_landscape', []):
164
- report_md += f"- **{item['Component']}**: {item['Findings']} ({item['Risk_Level']})\n"
165
-
166
- st.download_button("📥 Download Policy Report (.md)", report_md, f"Policy_Intel_{keyword}.md")
 
 
 
 
 
 
167
 
168
  # --- [Tab 2: Tactical Forensics] ---
169
  with tab2:
@@ -172,54 +219,61 @@ with tab2:
172
  if st.button("Analyze Violations"):
173
  with st.status(f"Scanning {target_geo} for Policy Breaches (20+ Sources)...", expanded=True):
174
  news = fetch_extensive_news(f"{target_geo} controversy OR scandal OR protest OR violence", geo=target_geo, limit=40, period="7d")
175
- news_titles = "\n".join([n['title'] for n in news])
176
-
177
- # 🚨 PROMPT UPDATE: 서머리에 '사건 개요' 필수 포함
178
- system_prompt = f"""
179
- Analyze Top 5 Risks in {target_geo} strictly through a 'Community Guidelines' lens.
180
 
181
- Summary MUST start with **"The Incident:"** (What happened) followed by **"Policy Analysis:"** (Why it violates rules).
182
-
183
- Return ONLY JSON:
184
- {{
185
- "risks": [
186
- {{
187
- "rank": 1, "title": "Event Title", "score": 90,
188
- "summary": "1. The Incident: ... \n2. Policy Analysis: ...",
189
- "forensic_grid": {{
190
- "Guideline_Breached": "e.g. Dangerous Organizations Policy",
191
- "Victim_Demographics": "e.g. Teenagers / Ethnic Minorities",
192
- "Societal_Impact": "e.g. Incitement to Violence",
193
- "Enforcement_Action": "e.g. Geo-blocking / Account Ban"
 
 
 
 
 
 
 
 
 
194
  }}
195
- }}
196
- ]
197
- }}
198
- """
199
- res = client.chat.completions.create(model=CURRENT_MODEL, messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": news_titles}], response_format={"type": "json_object"})
200
- report_data = json.loads(res.choices[0].message.content).get('risks', [])
201
-
202
- full_report_md = f"# NATIONAL T&S FORENSICS: {target_geo}\n\n"
203
-
204
- for i, r in enumerate(report_data):
205
- full_report_md += f"## {r.get('rank')}. {r.get('title')} (Risk: {r.get('score')})\n"
206
- full_report_md += f"{r.get('summary')}\n\n"
207
 
208
- with st.expander(f"🚩 RISK {r.get('rank')}: {r.get('title')} (Score: {r.get('score')})", expanded=True):
209
- c1, c2 = st.columns([1, 4])
210
- with c1:
211
- st.plotly_chart(render_gauge(r.get('score'), "Risk Index"), use_container_width=True, key=f"fg_{i}")
212
- with c2:
213
- st.markdown("**Incident & Policy Analysis:**")
214
- st.markdown(r.get('summary').replace(". ", ".\n\n"))
215
 
216
- st.table(pd.DataFrame(r.get('forensic_grid', {}).items(), columns=["T&S Component", "Assessment"]))
217
-
218
- st.download_button("📥 Download Forensic Report (.md)", full_report_md, f"T&S_Forensics_{target_geo}.md")
219
-
220
- st.divider()
221
- st.caption(f"Evidence Base: {len(news)} articles")
222
- st.dataframe(pd.DataFrame(news)[['title', 'source']], use_container_width=True)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
223
 
224
  # --- [Tab 3: Velocity] ---
225
  with tab3:
 
1
  import streamlit as st
2
  import feedparser
3
  import urllib.parse
4
+ from groq import Groq, RateLimitError # 에러 핸들링을 위해 추가
5
  import pandas as pd
6
  import json
7
  import os
 
23
  st.stop()
24
 
25
  client = Groq(api_key=api_key)
26
+
27
+ # 🚨 [NEW] Robust Model List (Fallback Strategy)
28
+ # 1순위 모델이 막히면 2순위, 3순위로 자동 전환하여 분석을 수행합니다.
29
+ MODEL_HIERARCHY = [
30
+ "llama-3.3-70b-versatile", # Best Quality
31
+ "llama3-70b-8192", # High Quality
32
+ "mixtral-8x7b-32768", # Good Alternative
33
+ "llama3-8b-8192" # Fast & Light (Emergency Backup)
34
+ ]
35
 
36
  # --- 3. Configuration ---
37
  REGION_MAP = {
 
42
  }
43
 
44
  # --- 4. Core Functions ---
45
+ # [NEW] Smart AI Request Function (Auto-Switching)
46
+ def get_ai_response(system_msg, user_content, json_mode=True):
47
+ for model in MODEL_HIERARCHY:
48
+ try:
49
+ kwargs = {
50
+ "model": model,
51
+ "messages": [{"role": "system", "content": system_msg}, {"role": "user", "content": user_content}],
52
+ }
53
+ if json_mode:
54
+ kwargs["response_format"] = {"type": "json_object"}
55
+
56
+ res = client.chat.completions.create(**kwargs)
57
+ return res # 성공하면 반환
58
+ except RateLimitError:
59
+ # 한도 초과 시 다음 모델로 조용히 넘어감
60
+ continue
61
+ except Exception as e:
62
+ # 다른 에러는 즉시 보고
63
+ st.error(f"⚠️ Error with model {model}: {e}")
64
+ return None
65
+
66
+ # 모든 모델이 실패했을 경우
67
+ st.error("🚨 All AI models are currently rate-limited or unavailable. Please try again in a few minutes.")
68
+ return None
69
+
70
  def fetch_extensive_news(query, geo="US", limit=60, period="7d"):
71
  time_filter = f" when:{period}"
72
  encoded = urllib.parse.quote(query + time_filter)
 
92
  fig.update_layout(height=180, margin=dict(l=10, r=10, t=40, b=10), paper_bgcolor="rgba(0,0,0,0)")
93
  return fig
94
 
95
+ # --- 5. 🚨 GLOBAL WAR ROOM ---
96
  st.title("🚨 OMNI-WATCH: T&S WAR ROOM")
97
  st.caption(f"Policy Risk Monitoring: {datetime.now().strftime('%H:%M:%S')} UTC")
98
 
 
102
  if 'hot_issues' not in st.session_state:
103
  with st.spinner("Scanning for Community Standard Violations..."):
104
  raw_news = fetch_extensive_news("violence OR hate speech OR disinformation OR scandal OR protest", limit=40, period="12h")
 
 
 
 
 
 
 
 
 
105
 
106
+ if not raw_news:
107
+ st.warning("No critical incidents in the last 12h.")
108
+ st.session_state.hot_issues = []
109
+ else:
110
+ news_context = "\n".join([f"Event: {n['title']}" for n in raw_news])
111
+
112
+ system_msg = """
113
+ Identify TOP 3 incidents with highest 'Trust & Safety' risk.
114
+ The 'summary' array MUST follow this order:
115
+ 1. "Event: [Brief summary]"
116
+ 2. "Violation: [Specific Guideline breached]"
117
+ 3. "Risk: [Offline harm]"
118
+
119
+ Return ONLY JSON:
120
+ {"issues": [{"title": "Short Title", "score": 85, "summary": ["Event: ...", "Violation: ...", "Risk: ..."], "link": ".."}]}
121
+ """
122
+
123
+ # [NEW] Use Smart Wrapper
124
+ res = get_ai_response(system_msg, news_context)
125
+ if res:
126
+ try:
127
+ st.session_state.hot_issues = json.loads(res.choices[0].message.content).get('issues', [])[:3]
128
+ except:
129
+ st.session_state.hot_issues = []
130
 
131
  if st.session_state.hot_issues:
132
  cols = st.columns(3)
 
155
  for geo in REGION_MAP[reg][:2]:
156
  all_news.extend(fetch_extensive_news(keyword, geo=geo, limit=6, period="7d"))
157
 
158
+ if not all_news:
159
+ st.error("No news found for this keyword.")
160
+ else:
161
+ news_summary = "\n".join([n['title'] for n in all_news[:45]])
162
+
163
+ global_prompt = f"""
164
+ Analyze '{keyword}' focusing on 'Community Guidelines' and 'Social Impact'.
165
+
166
+ 1. Executive Summary: Start with a clear **Event Summary** of what happened. Then, analyze the systemic policy risks and societal harm. (300+ words).
167
+ 2. Risk Landscape: Map findings to specific violations.
168
+
169
+ Return ONLY JSON:
170
+ {{
171
+ "executive_summary": "Start with [The Event], then move to [Policy Analysis].",
172
+ "risk_landscape": [
173
+ {{"Component": "Primary Violation", "Findings": "...", "Risk_Level": "High"}},
174
+ {{"Component": "Vulnerable Target", "Findings": "...", "Risk_Level": "Critical"}},
175
+ {{"Component": "Offline Harm", "Findings": "...", "Risk_Level": "High"}},
176
+ {{"Component": "Enforcement Gap", "Findings": "...", "Risk_Level": "Medium"}}
177
+ ],
178
+ "platform_intelligence": [
179
+ {{"Platform": "TikTok", "Assessment": "...", "Strategy": "..."}},
180
+ {{"Platform": "YouTube", "Assessment": "...", "Strategy": "..."}},
181
+ {{"Platform": "Meta", "Assessment": "...", "Strategy": "..."}},
182
+ {{"Platform": "X", "Assessment": "...", "Strategy": "..."}}
183
+ ],
184
+ "strategic_conclusion": "Final Trust & Safety recommendation."
185
+ }}
186
+ """
187
+
188
+ # [NEW] Use Smart Wrapper
189
+ res = get_ai_response(global_prompt, news_summary)
190
+
191
+ if res:
192
+ g_data = json.loads(res.choices[0].message.content)
193
+
194
+ with st.container(border=True):
195
+ st.subheader("1. Policy Impact Executive Summary")
196
+ st.markdown(g_data.get('executive_summary').replace(". ", ".\n\n"))
197
+
198
+ st.subheader("2. Guideline Violation Matrix")
199
+ st.dataframe(pd.DataFrame(g_data.get('risk_landscape')), hide_index=True, use_container_width=True)
200
+
201
+ st.subheader("3. Platform Enforcement Strategy")
202
+ st.table(pd.DataFrame(g_data.get('platform_intelligence')))
203
+
204
+ st.success(f"**T&S Recommendation:** {g_data.get('strategic_conclusion')}")
205
+
206
+ # DOWNLOAD
207
+ report_md = f"# OMNI-WATCH POLICY REPORT: {keyword.upper()}\n\n"
208
+ report_md += f"## 1. EXECUTIVE SUMMARY\n{g_data.get('executive_summary')}\n\n"
209
+ report_md += "## 2. VIOLATION LANDSCAPE\n"
210
+ for item in g_data.get('risk_landscape', []):
211
+ report_md += f"- **{item['Component']}**: {item['Findings']} ({item['Risk_Level']})\n"
212
+
213
+ st.download_button("📥 Download Policy Report (.md)", report_md, f"Policy_Intel_{keyword}.md")
214
 
215
  # --- [Tab 2: Tactical Forensics] ---
216
  with tab2:
 
219
  if st.button("Analyze Violations"):
220
  with st.status(f"Scanning {target_geo} for Policy Breaches (20+ Sources)...", expanded=True):
221
  news = fetch_extensive_news(f"{target_geo} controversy OR scandal OR protest OR violence", geo=target_geo, limit=40, period="7d")
 
 
 
 
 
222
 
223
+ if not news:
224
+ st.error(f"No recent controversy news found for {target_geo}.")
225
+ else:
226
+ news_titles = "\n".join([n['title'] for n in news])
227
+
228
+ system_prompt = f"""
229
+ Analyze Top 5 Risks in {target_geo} strictly through a 'Community Guidelines' lens.
230
+
231
+ Summary MUST start with **"The Incident:"** (What happened) followed by **"Policy Analysis:"** (Why it violates rules).
232
+
233
+ Return ONLY JSON:
234
+ {{
235
+ "risks": [
236
+ {{
237
+ "rank": 1, "title": "Event Title", "score": 90,
238
+ "summary": "1. The Incident: ... \n2. Policy Analysis: ...",
239
+ "forensic_grid": {{
240
+ "Guideline_Breached": "e.g. Dangerous Organizations Policy",
241
+ "Victim_Demographics": "e.g. Teenagers / Ethnic Minorities",
242
+ "Societal_Impact": "e.g. Incitement to Violence",
243
+ "Enforcement_Action": "e.g. Geo-blocking / Account Ban"
244
+ }}
245
  }}
246
+ ]
247
+ }}
248
+ """
249
+
250
+ # [NEW] Use Smart Wrapper
251
+ res = get_ai_response(system_prompt, news_titles)
 
 
 
 
 
 
252
 
253
+ if res:
254
+ report_data = json.loads(res.choices[0].message.content).get('risks', [])
 
 
 
 
 
255
 
256
+ full_report_md = f"# NATIONAL T&S FORENSICS: {target_geo}\n\n"
257
+
258
+ for i, r in enumerate(report_data):
259
+ full_report_md += f"## {r.get('rank')}. {r.get('title')} (Risk: {r.get('score')})\n"
260
+ full_report_md += f"{r.get('summary')}\n\n"
261
+
262
+ with st.expander(f"🚩 RISK {r.get('rank')}: {r.get('title')} (Score: {r.get('score')})", expanded=True):
263
+ c1, c2 = st.columns([1, 4])
264
+ with c1:
265
+ st.plotly_chart(render_gauge(r.get('score'), "Risk Index"), use_container_width=True, key=f"fg_{i}")
266
+ with c2:
267
+ st.markdown("**Incident & Policy Analysis:**")
268
+ st.markdown(r.get('summary').replace(". ", ".\n\n"))
269
+
270
+ st.table(pd.DataFrame(r.get('forensic_grid', {}).items(), columns=["T&S Component", "Assessment"]))
271
+
272
+ st.download_button("📥 Download Forensic Report (.md)", full_report_md, f"T&S_Forensics_{target_geo}.md")
273
+
274
+ st.divider()
275
+ st.caption(f"Evidence Base: {len(news)} articles")
276
+ st.dataframe(pd.DataFrame(news)[['title', 'source']], use_container_width=True)
277
 
278
  # --- [Tab 3: Velocity] ---
279
  with tab3: