Sync from GitHub via hub-sync

content_pipeline.py

@@ -181,10 +181,15 @@ def generate_deck_from_images(images: list, n: int = 12) -> list[Card]:
             "  answer     — a concise reference answer (1-3 sentences)\n"
             "  topic      — a short tag naming the concept (e.g. 'Cell Biology')\n"
             "  difficulty — integer: 1 (recall), 2 (application), 3 (synthesis)\n\n"
+            "Output format example — return ONE array of OBJECTS exactly like this "
+            "(not an array of strings), no other text:\n"
+            '[{"question": "What does X do?", "answer": "X does Y.", "topic": "Topic A", '
+            '"difficulty": 1}, {"question": "Why does Z occur?", "answer": "Because ...", '
+            '"topic": "Topic A", "difficulty": 2}]\n\n'
             "No prose, no markdown fences, no text outside the JSON array."
         )},
         {"role": "user", "content": list(images) + [
-            "Generate quiz questions from these document page images."]},
+            "Generate the JSON array of quiz questions from these document page images."]},
     ]
     # Same defensive strict-JSON + repair pass as the text path; images ride along
     # in the message content and llm.chat() routes them to the vision model.
@@ -264,10 +269,15 @@ def _cards_from_chunk(chunk: str) -> list[Card]:
             "  topic      — a short tag naming the concept (e.g. 'Cell Biology')\n"
             "  difficulty — integer: 1 (direct recall), 2 (application/explanation), "
             "3 (synthesis/comparison)\n\n"
+            "Output format example — return ONE array of OBJECTS exactly like this "
+            "(not an array of strings), no other text:\n"
+            '[{"question": "What does X do?", "answer": "X does Y.", "topic": "Topic A", '
+            '"difficulty": 1}, {"question": "Why does Z occur?", "answer": "Because ...", '
+            '"topic": "Topic A", "difficulty": 2}]\n\n'
             "Mix all three difficulty levels across the questions. "
             "No prose, no markdown fences, no explanation outside the JSON array."
         )},
-        {"role": "user", "content": f"Generate quiz questions from this passage:\n\n{chunk}"},
+        {"role": "user", "content": f"Passage:\n{chunk}\n\nGenerate the JSON array now."},
     ]
     # Strict-JSON with one repair pass: chat_json feeds a malformed reply back to
     # the model demanding clean JSON before giving up. A chunk that still won't
@@ -354,7 +364,9 @@ def regenerate(card: Card, direction: str) -> Card:
             "You rewrite a single quiz question to a different difficulty while "
             "keeping the SAME underlying concept. Stay grounded in the passage — "
             "do not introduce facts that are not in it. "
-            "Return ONLY a JSON object with keys: question, answer, topic."
+            "Return ONLY a JSON object with keys: question, answer, topic. "
+            "Example (return ONE object exactly like this, no other text):\n"
+            '{"question": "What does X do?", "answer": "X does Y.", "topic": "Topic A"}'
         )},
         {"role": "user", "content": (
             f"Passage:\n{card['source_chunk']}\n\n"

learning_engine.py

@@ -71,7 +71,10 @@ def grade_answer(card: Card, user_answer: str) -> GradeResult:
          "You grade a student's answer against a reference answer. "
          "Return ONLY a JSON object with keys: "
          "score (integer 0-5), explanation (string for the student), "
-         "missed_concept (short string naming what they got wrong, or \"\")."},
+         "missed_concept (short string naming what they got wrong, or \"\"). "
+         "Example (return ONE object exactly like this, no other text):\n"
+         '{"score": 3, "explanation": "You captured the main idea but missed a '
+         'detail.", "missed_concept": "the specific detail"}'},
         {"role": "user", "content":
          f"Question: {card['question']}\nReference answer: {card['answer']}\n"
          f"Student answer: {user_answer}\nGrade it."},
@@ -160,7 +163,9 @@ def generate_followups(card: Card, grade: GradeResult, n: int = 2) -> list[Card]
     messages = [
         {"role": "system", "content":
          "The student missed a concept. Generate follow-up quiz questions that "
-         "drill it. Return ONLY a JSON array with keys: question, answer, topic."},
+         "drill it. Return ONLY a JSON array of OBJECTS with keys: question, answer, "
+         "topic. Example (return ONE array exactly like this, no other text):\n"
+         '[{"question": "What is X?", "answer": "X is Y.", "topic": "Topic A"}]'},
         {"role": "user", "content":
          f"Original question: {card['question']}\n"
          f"Missed concept: {grade['missed_concept']}\n"

llm.py

@@ -192,12 +192,14 @@ def _chat_vision(messages: list[dict], max_tokens: int) -> str:
             inputs[k] = v.to(dtype=getattr(torch, _resolve_dtype_name()))
 
     with torch.no_grad():
+        # Greedy decoding (do_sample=False): every caller wants a strict JSON
+        # object/array, and greedy is markedly more reliable at that than sampling
+        # for MiniCPM-V — verified on GPU. enable_thinking is already False so the
+        # tight token budget goes to the answer, not a <think> preamble.
         out = _model.generate(
             **inputs,
             max_new_tokens=max_tokens,
-            do_sample=True,
-            temperature=0.7,
-            top_p=0.9,
+            do_sample=False,
             downsample_mode="16x",
         )
     gen = out[0][inputs["input_ids"].shape[1]:]