"""OCR, grading, and variant generation prompts""" import json import base64 from app.services.llm_clients import get_vision_client, get_deepseek_client OCR_PROMPT = """You are an expert at recognizing handwritten answers. Analyze this photo of a student's handwritten answer and extract the text and mathematical formulas. Requirements: - Faithfully extract what the student wrote, do not modify or correct - Use LaTeX format for math formulas (e.g. $x^2 + 1$) - If there are multiple steps, list them in original order - If some handwriting is unclear, mark with [unclear] Return only the extracted text, no additional explanation.""" GRADING_PROMPT = """You are an expert academic grader. Grade the following student answer. ALL output must be in English. Question info: - Number: {question_number} - Type: {question_type} - Question: {question_text} - Score: {score} Reference answer / solution: {reference_answer} Student answer: {student_answer} Grade and return JSON: {{ "is_correct": true/false, "score_given": 0-{score}, "feedback": " Step-by-step analysis of the student's answer, pointing out correct parts and errors, using KaTeX formulas ", "error_at_step": null or the step number where errors begin (integer) }} Grading rules: - MC / fill-blank: only correct if answer matches exactly - Long questions: give partial credit for correct steps even if the final answer is wrong - feedback in HTML format, supports KaTeX ($..$ inline, $$...$$ block) - Mark errors with
...
- Identify exactly which step the error starts""" VARIANT_PROMPT = """You are an expert exam question creator. Generate a similar but different variant question based on the original below. ALL output must be in English. Original question info: - Type: {question_type} - Question: {question_text} - Topics: {topics} - Difficulty: {difficulty} - Reference answer: {answer} Requirements: - Variant must test the same knowledge points at similar difficulty - Data/scenario/wording must differ — don't just change numbers - Must provide a complete correct answer Format requirements (CRITICAL): - All text in HTML format, absolutely NO markdown syntax - Code: 
...
, NOT ``` - Math: $...$ (inline) or $$...$$ (block), KaTeX compatible - Line breaks:
, paragraphs:

Return JSON: {{ "question_text": "HTML formatted variant question", "question_type": "{question_type}", "options": [MC only, format {{"label":"A","text":"..."}}, ...] or null, "correct_answer": "Correct answer (plain text)", "ai_hint": "HTML formatted hint that guides thinking WITHOUT giving the answer", "solution": "HTML formatted complete step-by-step solution" }}""" def ocr_photo(photo_bytes: bytes) -> str: """Gemini Vision OCR for handwritten answers""" client = get_vision_client() b64 = base64.b64encode(photo_bytes).decode("utf-8") resp = client.chat.completions.create( model="gemini-2.5-flash", messages=[ {"role": "system", "content": OCR_PROMPT}, {"role": "user", "content": [ {"type": "image_url", "image_url": { "url": f"data:image/jpeg;base64,{b64}", }}, ]}, ], temperature=0, max_tokens=2000, ) return resp.choices[0].message.content or "" def grade_answer(question: dict, student_answer: str) -> dict: """Qwen grades student answer""" reference = question.get("raw_answer_text") or question.get("solution") or "No reference answer" score = question.get("score") or "unknown" ds = get_deepseek_client() resp = ds.chat.completions.create( model="deepseek-chat", messages=[ {"role": "system", "content": GRADING_PROMPT.format( question_number=question["question_number"], question_type=question["question_type"], question_text=question["question_text"], score=score, reference_answer=reference, student_answer=student_answer, )}, ], temperature=0.2, response_format={"type": "json_object"}, ) return json.loads(resp.choices[0].message.content) def generate_variant(question: dict) -> dict: """Gemini generates a variant question""" answer = ( question.get("correct_option") or question.get("correct_answer") or question.get("raw_answer_text") or "N/A" ) ds = get_deepseek_client() resp = ds.chat.completions.create( model="deepseek-chat", messages=[ {"role": "system", "content": VARIANT_PROMPT.format( question_type=question["question_type"], question_text=question["question_text"], topics=", ".join(question.get("topics", [])), difficulty=question.get("difficulty", "medium"), answer=answer, )}, ], temperature=0.5, response_format={"type": "json_object"}, ) return json.loads(resp.choices[0].message.content)