cot

app.py

@@ -246,6 +246,7 @@ CHAIN_CSS = """
 
 .thinking { font-size: 13px; color: #888; padding: 8px 0; }
 .empty-hint { font-size: 12px; color: #ccc; padding: 6px 0; }
+.temporal-line { font-size: 11px; color: #666; margin-top: 8px; font-family: 'IBM Plex Mono', monospace; }
 
 .prompt-snippet {
     font-size: 11px; color: #888; line-height: 1.5;
@@ -289,108 +290,137 @@ def render_chain(s1_text, s2_text, s3_text, status="done"):
     s2_active = status in ("running2", "running3", "done")
     s3_active = status in ("running3", "done")
 
-    # ── Stage 1: extract bullet lines from free text ──────────────────────────
+    # ── Stage 1 ───────────────────────────────────────────────────────────────
     if status == "running1":
         s1_content = f'<div class="thinking">Extracting features {_dots()}</div>'
     elif s1_text:
         tags = []
+        # Parse LOCATION INVENTORY bullets: "- Name: N visits, description"
+        in_inventory = False
         for line in s1_text.splitlines():
-            line = line.strip().lstrip("-").strip()
-            # Keep lines that look like "Location: N visits, description"
-            if line and len(line) > 8 and len(line) < 80 and not line.endswith(":"):
-                tags.append(line)
-            if len(tags) >= 8:
+            line = line.strip()
+            if "LOCATION INVENTORY" in line.upper():
+                in_inventory = True
+                continue
+            if in_inventory:
+                if line.startswith("TEMPORAL") or line.startswith("SEQUENCE") or (line and not line.startswith("-") and not line.startswith("*") and len(line) > 40):
+                    break
+                if line.startswith("-"):
+                    # "- Name: N visits, type"  or  "- Name (N visits)"
+                    clean = line.lstrip("-").strip()
+                    # Shorten: keep "Name (N visits)" style
+                    m = re.match(r'(.+?):\s*(\d+)\s*visit', clean, re.IGNORECASE)
+                    if m:
+                        name = m.group(1).strip()
+                        n    = m.group(2)
+                        tags.append(f"{name} · {n}x")
+                    elif clean:
+                        tags.append(clean[:55])
+                    if len(tags) >= 8:
+                        break
+
+        # Fallback: also grab temporal summary line
+        temporal_line = ""
+        for line in s1_text.splitlines():
+            line = line.strip()
+            if "weekly distribution" in line.lower() or "weekday" in line.lower():
+                temporal_line = line.lstrip("-").strip()[:70]
                 break
-        s1_content = '<div class="tag-row">' + "".join(f'<span class="tag">{t}</span>' for t in tags) + '</div>'
+
+        tag_html = "".join(f'<span class="tag">{t}</span>' for t in tags)
+        temp_html = f'<div class="temporal-line">⏱ {temporal_line}</div>' if temporal_line else ""
+        s1_content = f'<div class="tag-row">{tag_html}</div>{temp_html}'
     else:
         s1_content = '<div class="empty-hint">Press ▶ to start</div>'
 
-    # ── Stage 2: extract numbered sections as key-value rows ──────────────────
+    # ── Stage 2 ───────────────────────────────────────────────────────────────
     KEYS = [
         ("SCHEDULE",  ["ROUTINE", "SCHEDULE"]),
         ("ECONOMIC",  ["ECONOMIC", "SPENDING", "FINANCIAL"]),
-        ("SOCIAL",    ["SOCIAL", "COMMUNITY"]),
-        ("LIFESTYLE", ["LIFESTYLE", "ACTIVITY", "LEISURE"]),
-        ("STABILITY", ["STABILITY", "CONSISTENCY", "PATTERN"]),
+        ("SOCIAL",    ["SOCIAL", "COMMUNITY", "LIFESTYLE"]),
+        ("STABILITY", ["STABILITY", "CONSISTENCY", "REGULARITY"]),
     ]
     if status == "running2":
         s2_content = f'<div class="thinking" style="color:#a06030">Analyzing behavior {_dots()}</div>'
     elif s2_text:
-        rows_html = ""
-        lines = s2_text.splitlines()
-        # Build a searchable blob per section by scanning numbered headings
+        # Parse numbered sections
         sections = {}
         current_key = None
-        current_lines = []
-        for line in lines:
+        current_bullets = []
+        for line in s2_text.splitlines():
             line = line.strip()
-            # Detect numbered headings like "1. ROUTINE & SCHEDULE ANALYSIS:"
-            heading_match = re.match(r'^\d+\.\s+(.+?):\s*$', line)
-            if heading_match:
+            m = re.match(r'^\d+\.\s+(.+?)(?:\s+ANALYSIS)?(?:\s+PATTERNS)?(?:\s+INDICATORS)?:\s*$', line, re.IGNORECASE)
+            if m:
                 if current_key:
-                    sections[current_key] = " ".join(current_lines)
-                current_key = heading_match.group(1).upper()
-                current_lines = []
+                    sections[current_key] = current_bullets
+                current_key = m.group(1).upper()
+                current_bullets = []
             elif current_key and line.startswith("-"):
-                current_lines.append(line.lstrip("-").strip())
+                bullet = line.lstrip("-").strip()
+                if bullet:
+                    current_bullets.append(bullet)
         if current_key:
-            sections[current_key] = " ".join(current_lines)
+            sections[current_key] = current_bullets
 
+        rows_html = ""
         for label, search_words in KEYS:
             val = "—"
-            for k, v in sections.items():
-                if any(w in k for w in search_words):
-                    # Take first sentence
-                    first = v.split(".")[0].strip() if v else ""
-                    val = first[:90] + ("..." if len(first) > 90 else "")
+            for k, bullets in sections.items():
+                if any(w in k for w in search_words) and bullets:
+                    # Take first bullet, truncate at 2 sentences
+                    text = bullets[0]
+                    sentences = re.split(r'(?<=[.!?])\s+', text)
+                    val = " ".join(sentences[:2])
+                    if len(val) > 100:
+                        val = val[:97] + "..."
                     break
             rows_html += f'<div class="bkey">{label}</div><div class="bval">{val}</div>'
         s2_content = f'<div class="behavior-row">{rows_html}</div>'
     else:
         s2_content = '<div class="empty-hint">Waiting...</div>'
 
-    # ── Stage 3: prediction ───────────────────────────────────────────────────
+    # ── Stage 3 ───────────────────────────────────────────────────────────────
     if status == "running3":
         s3_content = f'<div class="thinking" style="color:#c0392b">Inferring demographics {_dots()}</div>'
     elif s3_text:
-        pred = reasoning = alts = ""
-        for line in s3_text.splitlines():
-            line = line.strip()
+        pred = reasoning = ""
+        lines = s3_text.splitlines()
+        i = 0
+        while i < len(lines):
+            line = lines[i].strip()
             if line.startswith("INCOME_PREDICTION:"):
                 pred = line.replace("INCOME_PREDICTION:", "").strip()
             elif line.startswith("INCOME_REASONING:"):
                 reasoning = line.replace("INCOME_REASONING:", "").strip()
-            elif line.startswith("ALTERNATIVES:"):
-                alts = line.replace("ALTERNATIVES:", "").strip()
-        # If reasoning spans multiple lines, grab all of it
-        if not reasoning:
-            in_reasoning = False
-            reasoning_parts = []
-            for line in s3_text.splitlines():
-                line = line.strip()
-                if line.startswith("INCOME_REASONING:"):
-                    in_reasoning = True
-                    reasoning_parts.append(line.replace("INCOME_REASONING:", "").strip())
-                elif in_reasoning and line and not line.startswith("INCOME_") and not line.startswith("ALTERNATIVES"):
-                    reasoning_parts.append(line)
-                elif in_reasoning and (line.startswith("INCOME_") or line.startswith("ALTERNATIVES")):
-                    break
-            reasoning = " ".join(reasoning_parts).strip()
+                # Collect continuation lines until blank or next key
+                i += 1
+                while i < len(lines):
+                    nxt = lines[i].strip()
+                    if not nxt or nxt.startswith("INCOME_") or re.match(r'^\d+\.', nxt):
+                        break
+                    reasoning += " " + nxt
+                    i += 1
+                continue
+            i += 1
+
+        # Truncate reasoning to ~2 sentences
+        sentences = re.split(r'(?<=[.!?])\s+', reasoning.strip())
+        short_reasoning = " ".join(sentences[:2])
+        if len(short_reasoning) > 160:
+            short_reasoning = short_reasoning[:157] + "..."
 
-        alts_html = f'<div class="alternatives">Also possible: <span>{alts}</span></div>' if alts else ""
         s3_content = f"""
         <div class="pred-block">
             <div class="pred-label">Income Prediction</div>
             <div class="pred-value">{pred or "—"}</div>
-            <div class="reasoning-text">{reasoning}</div>
-            {alts_html}
+            <div class="reasoning-text">{short_reasoning}</div>
         </div>"""
     else:
         s3_content = '<div class="empty-hint">Waiting...</div>'
 
     PROMPT_SNIPPETS = {
         "s1": "You are an expert mobility analyst. Given the trajectory data below, extract: (1) LOCATION INVENTORY — list all POI categories visited and visit frequency; (2) TEMPORAL PATTERNS — weekly distribution, peak hours; (3) SEQUENCE — typical activity chains...",
-        "s2": "Based on the trajectory features identified: {Response 1}. Now analyze what these mobility patterns reveal about lifestyle: (1) SCHEDULE — work/activity routine type; (2) ECONOMIC — spending venue tiers; (3) SOCIAL — social engagement; (4) LIFESTYLE — activity diversity; (5) STABILITY — consistency of routine...",
+        "s2": "Based on the trajectory features identified: {Response 1}. Now analyze what these mobility patterns reveal about lifestyle: (1) SCHEDULE — work/activity routine type; (2) ECONOMIC — spending venue tiers; (3) SOCIAL — social engagement; (4) STABILITY — consistency of routine...",
         "s3": "Based on feature analysis {Response 1} and behavioral analysis {Response 2}, predict income level. Output — INCOME_PREDICTION: [range]; INCOME_REASONING: [detailed reasoning]...",
     }
 
@@ -430,8 +460,6 @@ def render_chain(s1_text, s2_text, s3_text, status="done"):
     return html
 
 
-# ── Map & demo ────────────────────────────────────────────────────────────────
-
 def build_map(agent_sp):
     agent_sp = agent_sp.reset_index(drop=True).copy()
     agent_sp["latitude"]  += np.random.uniform(-0.0003, 0.0003, len(agent_sp))