Spaces:
Paused
Paused
Commit History
-
Refactor to add answer grid & assoicate llm parsing
Yu Chenad5d4b6
-
Fix: Clarify default prompt — non-dash answers are wrong answers
chih.yikuan +1: Claude Opus 4.6 (1M context)Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>bb875d5 -
Simplify: send raw data to LLM, let user's prompt control interpretation
chih.yikuan +1: Claude Opus 4.6 (1M context)Remove all backend filtering logic (correct markers, teacher answer comparison). Now sends raw student answers + all questions + teacher answers (if any) directly into the prompt template. The user controls how data is interpreted via the prompt editor (e.g. "'-' means correct"). Default prompt includes data format instructions. Preview and generate endpoints both accept the user's prompt_template. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>34b22f3 -
Fix: Treat '-' and similar markers as correct answers in student data
chih.yikuan +1: Claude Opus 4.6 (1M context)Student answer data often uses '-', 'v', '✓' etc. to mark correct answers. filter_student_wrong_answers now skips these markers and only includes actual wrong answer text. Works with or without teacher answers uploaded. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>c99a18b -
Fix: Handle missing teacher answers — include all student answers instead
chih.yikuan +1: Claude Opus 4.6 (1M context)When 標準答案 is not uploaded, filter_student_wrong_answers now returns ALL student answers + ALL questions (instead of empty arrays), with a separate prompt template that asks the LLM to infer correctness. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>3efc9d0 -
Add final prompt preview tab with real student data substitution
chih.yikuan +1: Claude Opus 4.6 (1M context)DataPreviewModal now opens on "最終提示詞" tab first, showing the actual filled prompt for a selected student with {wrong_questions}, {wrong_answers} etc. replaced by real data. Users can switch students via dropdown to see each student's specific prompt. Textareas are resizable. Backend adds /preview-student-prompt endpoint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>5722fbf -
Redesign Step 2: per-student report generation with wrong-answer filtering
chih.yikuan +1: Claude Opus 4.6 (1M context)Step 2 now shows a student checklist instead of prompt editor. Users select which students to generate reports for. Backend filters each student's wrong answers and only sends the relevant questions + wrong answers to the LLM, dramatically reducing prompt size. Reports generate sequentially with live progress in Step 3's tabbed viewer. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>f002ae5
-
Fix: Add prompt size check and auto-truncate students to prevent empty LLM responses
chih.yikuan +1: Claude Opus 4.6 (1M context)When exam data is too large for the model's context window, the report generator now automatically truncates student data (50% → 25%) instead of failing. Also switches to compact JSON serialization to save ~20-30% tokens, and dynamically adjusts max_completion_tokens based on available space. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>30ff567
-
Redesign ClassLens v2: 3-step exam analysis workflow
chih.yikuan +1: Claude Opus 4.6Replace ChatKit-based chat UI with a structured 3-step workflow: 1. Upload & parse exam files (questions, student answers, teacher answers) via GPT Vision 2. Edit analysis prompts with save/load/share functionality 3. Generate and export HTML reports (PDF/PNG) - Add JWT email/password auth (replace Google OAuth) - Add file processor with multi-format support (PDF, images, text files) - Add image-to-text description for visual exam content - Add model selector (GPT-5.4, 5.3 Instant, 4o, etc.) - Add per-zone file upload with inline parsed data preview - Remove ChatKit SDK, Google Sheets, email service dependencies - Remove .env from tracking (contains API keys) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>bc45e54