Fix imports, remove dead code, and make token handling model-aware in chat_paper and chat_arxiv

[Copilot]

Several correctness and robustness issues in chat_paper.py and chat_arxiv.py caused inaccurate token counting, silent text truncation bugs, and hardcoded context limits that broke for GPT-4 and 16k models.

Changes

• Fix duplicate import os / multi-module import (chat_paper.py): import fitz, io, os shadowed the earlier import os; split into clean, separate stdlib and third-party import blocks.

• Remove unused images variable (both files): images = page.get_images() was assigned and immediately shadowed by the for … in page.get_images() loop — dead assignment removed.

• Model-aware max_token_num (both files): Replaced the hardcoded 4096 with a lookup dict so GPT-3.5-turbo-16k, GPT-4, GPT-4-32k, and GPT-4-turbo users get the correct context window. Unknown models fall back to 4096.

model_max_tokens = {
    'gpt-3.5-turbo': 4096,
    'gpt-3.5-turbo-16k': 16384,
    'gpt-4': 8192,
    'gpt-4-32k': 32768,
    'gpt-4-turbo': 128000,
}
self.max_token_num = model_max_tokens.get(self.chatgpt_model, 4096)

• Model-aware tiktoken encoding (both files): tiktoken.get_encoding("gpt2") gives wrong token counts for GPT-3.5/4 (which use cl100k_base). Replaced with tiktoken.encoding_for_model(self.chatgpt_model) with a cl100k_base fallback for unknown models.

• Bounds-check clip_text_index (both files, all three chat methods): The index calculation could yield 0 or negative when prompt token budget ≥ text length, producing an empty or inverted slice. Wrapped with max(1, …).

• Read model from config in chat_arxiv.py: The model was hardcoded to "gpt-3.5-turbo" in all three openai.ChatCompletion.create calls. Now reads CHATGPT_MODEL from apikey.ini via self.chatgpt_model, consistent with chat_paper.py.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.