Use bounded PDF reading and source-guided verification; honor requested dates and image coordinates

app.py

@@ -30,6 +30,7 @@ MODEL = os.getenv("GROQ_MODEL", "openai/gpt-oss-120b")
 INPUT_PATH = Path("/tmp/course-agent-private-inputs.json")
 CACHE_PATH = Path("/tmp/course-agent-owner-records.json")
 MAX_BYTES = 10 * 1024 * 1024
+MAX_PDF_BYTES = 40 * 1024 * 1024
 USER_AGENT = "HF-Course-Agent/1.1 (https://huggingface.co/spaces/yongchoooon/Final_Assignment_Template)"
 
 
@@ -71,8 +72,9 @@ def download(url):
             chunks, size = [], 0
             for chunk in response.iter_content(65536):
                 size += len(chunk)
-                if size > MAX_BYTES:
-                    raise ValueError("File exceeds the 10 MB reading limit.")
+                limit = MAX_PDF_BYTES if (chunks and chunks[0].startswith(b"%PDF")) or (not chunks and chunk.startswith(b"%PDF")) else MAX_BYTES
+                if size > limit:
+                    raise ValueError("File exceeds the bounded reading limit.")
                 chunks.append(chunk)
             return b"".join(chunks)
     raise ValueError("Too many redirects.")
@@ -349,9 +351,10 @@ def attachment_tool(item):
                     {"type": "text", "text": "Analyze this attached image for the following question. "
                      "Describe precisely what is visible and relevant; distinguish observations from deductions. "
                      "For a chess board, give exact piece positions and a complete FEN, checking square coordinates and orientation. "
+                     "Read edge labels; do not assume White is at the bottom. Use the printed files and ranks. "
                      "Do not guess a winning move: a separate chess engine will verify the FEN. " + item["question"]},
                     {"type": "image_url", "image_url": {"url": "data:" + mime + ";base64," + base64.b64encode(content).decode()}}]}],
-                max_tokens=2048)
+                max_tokens=2048, reasoning_effort="none", reasoning_format="hidden")
             if response.choices[0].finish_reason == "length":
                 raise ValueError("Image analysis was truncated.")
             return "Image analysis: " + (response.choices[0].message.content or "")
@@ -393,10 +396,13 @@ def solve(item, trace=None, progress=None):
         "Use primary sources. Never treat a shopping/search-results page that merely repeats a question "
         "as evidence. For named-source questions, read and confirm the claim in the specified source "
         "before answering. A failed page download does not provide evidence. For English Wikipedia use read_wikipedia. If a historical year is "
-        "specified, set as_of to December 31 of that year. Select the relevant article section, "
+        "specified, use its exact cutoff: last day of the named month if a month is given, "
+        "December 31 only when just a year is given. Select the relevant article section, "
         "and paginate if more text is available. Read an attachment when present. "
         "Check every constraint and exclusion before answering. Culinary categories are not botanical "
-        "categories. For ranking/counts verify the complete relevant table, not a search snippet. "
+        "categories. Classify the actual edible plant organ (leaf, stem, root, flower, fruit, seed). "
+        "Exclude botanical fruits even if used as culinary vegetables; do not overlook edible herb leaves. "
+        "For ranking/counts verify the complete relevant table, not a search snippet. "
         "For chess positions extract the FEN from the image, then verify the move with analyse_chess_position. "
         "Prefer narrower entity searches over quoting an entire long question. "
         "Do not repeat identical failed requests. Never claim to have read unsupported media. "
@@ -631,7 +637,7 @@ def import_answers(text, profile: gr.OAuthProfile | None):
     return f"답안 {len(records)}개 복원 완료. 아직 제출하지 않았습니다.", records, table, backup
 
 
-def evaluate(limit, retry_ids, records, profile: gr.OAuthProfile | None, progress=gr.Progress()):
+def evaluate(limit, retry_ids, source_hints, records, profile: gr.OAuthProfile | None, progress=gr.Progress()):
     authorize(profile)
     records = dict(records or {})
     try:
@@ -653,7 +659,13 @@ def evaluate(limit, retry_ids, records, profile: gr.OAuthProfile | None, progres
             table, backup = export_answers(records)
             yield f"{index}/{len(pending)} 풀이 중. 무료 한도에 맞춰 호출 간 대기할 수 있습니다.", records, table, backup
             try:
-                answer = solve(item, trace=trace, progress=progress)
+                solving_item = item
+                if source_hints and source_hints.strip():
+                    urls = re.findall(r"https?://[^\s]+", source_hints)[:6]
+                    if not urls:
+                        raise ValueError("참고 자료는 공개 원문 URL을 입력해 주세요.")
+                    solving_item = dict(item, question=item["question"] + "\nAdditional primary-source URLs to inspect first: " + "\n".join(urls))
+                answer = solve(solving_item, trace=trace, progress=progress)
                 records[item["task_id"]] = {"question": item["question"], "answer": answer, "trace": trace}
             except Exception as error:
                 message = error_message(error)
@@ -700,7 +712,7 @@ def submit(records, profile: gr.OAuthProfile | None):
 
 
 with gr.Blocks() as demo:
-    gr.Markdown("# Unit 4 · 나의 AI 에이전트 · v7")
+    gr.Markdown("# Unit 4 · 나의 AI 에이전트 · v8")
     gr.Markdown("로그인 → 연결 확인 → 한 문제 시험 → 전체 풀이 → 답안 확인 → 제출")
     gr.LoginButton()
     records = gr.State({})
@@ -717,13 +729,14 @@ with gr.Blocks() as demo:
         restore_json = gr.Textbox(label="기존 답안 JSON", lines=3)
         import_button = gr.Button("JSON 답안 복원")
         retry_ids = gr.Textbox(label="다시 풀 문제 ID (선택)", info="비워 두면 미완료 문제만 풉니다.")
+        source_hints = gr.Textbox(label="참고할 자료 URL (선택)", info="원문 링크를 알고 있을 때 입력하면 해당 자료부터 확인합니다.")
     with gr.Accordion("소유자 전용 첨부자료 복원", open=False):
         input_json = gr.Textbox(label="첨부자료 JSON", lines=2)
         input_button = gr.Button("원본 첨부자료 복원")
     submit_button = gr.Button("3. 현재 답안을 공식 리더보드에 제출")
     gr.Markdown("검색·위키백과·계산·Python·PDF·엑셀·음성·이미지를 지원합니다. 영상의 장면 분석은 아직 지원하지 않습니다.")
     connection.click(test_connection, outputs=status, concurrency_id="agent", concurrency_limit=1)
-    run_event = run.click(evaluate, inputs=[limit, retry_ids, records], outputs=[status, records, results, backup],
+    run_event = run.click(evaluate, inputs=[limit, retry_ids, source_hints, records], outputs=[status, records, results, backup],
               concurrency_id="agent", concurrency_limit=1)
     stop.click(fn=None, cancels=[run_event], queue=False)
     input_button.click(import_inputs, inputs=input_json, outputs=status, concurrency_id="inputs", concurrency_limit=1)