Update app.py

app.py

@@ -251,15 +251,36 @@ def _tts_to_b64(text: str) -> str:
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-#  Prompt builder — trimmed for 150-token output budget (Qwen3-0.6B, CPU)
+#  Prompt builder — trimmed for 180-token output budget (Qwen3-0.6B, CPU)
+#
+#  Key design: only the ACTIVE topic/goal is passed to stage_updates context.
+#  Showing all topics caused the model to update every entry, blowing the
+#  token budget and truncating the JSON.
 # ─────────────────────────────────────────────────────────────────────────────
+_STAGES = ("teach", "re_teach", "show_and_tell", "assess")
+
+def _find_active_topic(current_learning: list) -> tuple:
+    """Return (topic_name, goal_name, stage) for the first incomplete objective."""
+    for item in current_learning:
+        topic = item.get("topic", "")
+        for obj in item.get("learning_objectives", []):
+            goal = obj.get("goal", "")
+            for stage in _STAGES:
+                if obj.get(stage, "Not_Complete") != "complete":
+                    return topic, goal, stage
+    return "", "", "teach"   # all complete — nothing active
+
+
 def _build_system_prompt(lp: dict, rag_chunks: List[str]) -> str:
     persona          = lp.get("teacher_persona", "A friendly and patient teacher")
     student          = lp.get("student_name", "Student")
-    chat_history     = lp.get("chat_history", [])[-2:]      # last 2 turns only
-    scratchpad       = lp.get("scratchpad", [])[-1:]         # last 1 entry only
+    chat_history     = lp.get("chat_history", [])[-2:]   # last 2 turns only
+    scratchpad       = lp.get("scratchpad",   [])[-1:]   # last 1 entry only
     current_learning = lp.get("assessment_stages", {}).get("current_learning", [])
 
+    # ── Find the single active topic/goal to teach right now ─────────────────
+    active_topic, active_goal, active_stage = _find_active_topic(current_learning)
+
     history_block = "\n".join(
         f'S: {h.get("user_input","")}\nT: {h.get("system_output","")}'
         for h in chat_history
@@ -271,13 +292,19 @@ def _build_system_prompt(lp: dict, rag_chunks: List[str]) -> str:
     ) or "Empty."
 
     rag_block = "\n---\n".join(rag_chunks) if rag_chunks else "No relevant content found."
-    cl_block  = json.dumps(current_learning, indent=2) if current_learning else "[]"
+
+    # Pass only the active topic/goal — not the whole list — to keep output short
+    active_block = (
+        f'Topic: "{active_topic}"\nGoal: "{active_goal}"\nCurrent stage: {active_stage}'
+        if active_topic else "All objectives complete."
+    )
 
     return f"""You are {persona} teaching {student}, aged 6–12. Use simple English. Be warm and brief.
 
 STUDENT: {student}
-LEARNING OBJECTIVES:
-{cl_block}
+
+ACTIVE OBJECTIVE (teach this now):
+{active_block}
 
 KNOWLEDGE BASE:
 {rag_block}
@@ -288,66 +315,92 @@ RECENT CHAT:
 NOTES:
 {scratch_block}
 
-TASK: Classify intent, respond to the student, return ONLY valid JSON. Keep "response" under 60 words.
+TASK: Classify intent, respond to the student, return ONLY valid JSON. Keep "response" under 50 words.
 
 INTENT RULES:
-"block" — rude/inappropriate message. First time: redirect kindly. Repeat: end gently.
-"questions" — off-topic question. Answer briefly from KB if found, then redirect.
-"curriculum" — on-topic. Follow stages in order: teach → re_teach → show_and_tell → assess.
-  teach: explain using KB. re_teach: ask one check question; re-explain if wrong.
-  show_and_tell: ask a similar question. assess: pass=complete, fail=Not_Complete (retry).
-"chitchat" — casual talk. Respond warmly, then bring up learning topic.
-
-OUTPUT — return ONLY this JSON:
+"block" — rude/inappropriate. Redirect kindly (first time) or end gently (repeat).
+"questions" — off-topic. Answer briefly from KB, then redirect.
+"curriculum" — on-topic. Follow: teach → re_teach → show_and_tell → assess.
+"chitchat" — casual. Respond warmly, bring up active topic.
+
+OUTPUT — return ONLY this JSON (stage_updates: EXACTLY 1 entry for the active objective only):
 {{
   "intent": "<block|questions|curriculum|chitchat>",
-  "response": "<reply to student, max 60 words>",
-  "stage_updates": [{{"topic":"<topic>","goal":"<goal>","teach":"<complete|Not_Complete>","re_teach":"<complete|Not_Complete>","show_and_tell":"<complete|Not_Complete>","assess":"<complete|Not_Complete>"}}],
-  "thought": "<brief internal reasoning>",
+  "response": "<reply, max 50 words>",
+  "stage_updates": [{{"topic":"{active_topic}","goal":"{active_goal}","teach":"<complete|Not_Complete>","re_teach":"<complete|Not_Complete>","show_and_tell":"<complete|Not_Complete>","assess":"<complete|Not_Complete>"}}],
+  "thought": "<one sentence>",
   "action": "<teach|re_teach|show_and_tell|assess|answer|redirect|discourage|end|chitchat>",
-  "observation": "<brief student observation>"
+  "observation": "<one sentence>"
 }}\
 """
 
 
 # ─────────────────────────────────────────────────────────────────────────────
-#  JSON parser  (robust — handles markdown fences, partial JSON, etc.)
+#  JSON parser — layered extraction, regex-anchored on "intent" key.
+#
+#  Layer 0 : strip any <think>…</think> block (Qwen3 safety fallback).
+#  Layer 1 : strip markdown ```json … ``` fences.
+#  Layer 2 : direct json.loads on the cleaned text.
+#  Layer 3 : regex — walk every '{' left-to-right; skip those that don't
+#             contain "intent":; try every '}' right-to-left until a valid
+#             JSON object with "intent" key parses successfully.
+#  Layer 4 : broad regex — outermost { … } regardless of content.
+#  Layer 5 : fallback dict with raw text as the response field.
 # ─────────────────────────────────────────────────────────────────────────────
+import re as _re
+
 def _parse_llm_output(raw: str) -> dict:
-    text = raw.strip()
-
-    # Strip Qwen3 <think>…</think> block if present (safety fallback)
-    if text.startswith("<think>"):
-        end = text.find("</think>")
-        text = text[end + len("</think>"):].strip() if end != -1 else text
-
-    if "```" in text:
-        for part in text.split("```"):
-            part = part.strip()
-            if part.startswith("json"):
-                part = part[4:].strip()
-            try:
-                return json.loads(part)
-            except json.JSONDecodeError:
-                continue
+    # ── Layer 0: strip Qwen3 <think>…</think> block ──────────────────────────
+    text = _re.sub(r"<think>.*?</think>", "", raw, flags=_re.DOTALL).strip()
+
+    # ── Layer 1: strip markdown fences ───────────────────────────────────────
+    fence_match = _re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, _re.DOTALL)
+    if fence_match:
+        try:
+            return json.loads(fence_match.group(1))
+        except json.JSONDecodeError:
+            pass
 
+    # ── Layer 2: direct parse ─────────────────────────────────────────────────
     try:
         return json.loads(text)
     except json.JSONDecodeError:
         pass
 
-    start = text.find("{")
-    end   = text.rfind("}") + 1
-    if start != -1 and end > start:
+    # ── Layer 3: intent-anchored brace scan ───────────────────────────────────
+    intent_pat   = _re.compile(r'"intent"\s*:', _re.DOTALL)
+    brace_opens  = [m.start() for m in _re.finditer(r'\{', text)]
+    brace_closes = [m.end()   for m in _re.finditer(r'\}', text)]
+
+    for open_pos in brace_opens:
+        region = text[open_pos:]
+        if not intent_pat.search(region):
+            continue                          # no "intent": inside this brace
+        for close_pos in reversed(brace_closes):
+            if close_pos <= open_pos:
+                break
+            candidate = text[open_pos:close_pos]
+            try:
+                parsed = json.loads(candidate)
+                if "intent" in parsed:
+                    log.info("JSON extracted via intent-anchored regex.")
+                    return parsed
+            except json.JSONDecodeError:
+                continue
+
+    # ── Layer 4: outermost { … } fallback ────────────────────────────────────
+    broad = _re.search(r'\{.*\}', text, _re.DOTALL)
+    if broad:
         try:
-            return json.loads(text[start:end])
+            return json.loads(broad.group())
         except json.JSONDecodeError:
             pass
 
-    log.warning("Could not parse JSON from model output. Using raw text as response.")
+    # ── Layer 5: give up ─────────────────────────────────────────────────────
+    log.warning("Could not parse JSON from model output. Raw: %.200s", raw)
     return {
         "intent":        "questions",
-        "response":      raw,
+        "response":      text or raw,
         "stage_updates": [],
         "thought":       "",
         "action":        "answer",
@@ -620,4 +673,4 @@ if __name__ == "__main__":
         port=7860,
         log_level="info",
         workers=1,          # Single worker — shared in-memory model object
-    )
+    )