Bot for Extracting & Highlighting Outstanding Literature, Documentation, Evaluations & Reviews (BEHOLDER).

A python script to automatically retrieve Computer Science papers with predetermined watch terms using Semantic Scholar API, then generating an AI podcast using google cloud API. This python package requires the following packages to be importable (yes I should have exported a conda env, maybe later :p)

import os
import sys
import io
import requests
import pdfkit   # Can be very capricious depending on your operating system. Check if it works correctly!
import time
import enum
import json
import argparse
from pathvalidate import sanitize_filename
from datetime import datetime, timedelta, date
from dotenv import load_dotenv
from google import genai
from google.cloud import texttospeech
from pydantic import BaseModel
from pydub import AudioSegment   # Requires ffmpeg!
from tqdm import tqdm

The following needs to be done for installation (outside of installing python packages):

• make a .env file containing a value for G_API_KEY. The value should be a Google Cloud API key that's autorised for the following APIs: Generative Language API, Cloud Text-to-Speech API
• install FFmpeg if not already done

You can customize your paper search & podcast by doing the following:

• Change watch_terms.txt to your search terms (more search terms = more papers)
• Change script_prompt.txt to your liking. I recommend not editing it too drastically, but you can have fun with it. The {DAYS_BACK} and {INTERESTS} tags will be automatically replaced when running the program. You can place them wherever you want. Note that the {INTERESTS} are very important to force the script generation to focus on what actually interests you. If none are specified with the --interests option, it will use whatever your search terms are as your of area interests (which can work really fine if your search terms aren't too broad). I also wouldn't recommend changing the "output format" section of the prompt (besides names). Note that you can also point to different watch_terms.txt and script_prompt.txt files using the --watch_terms & --script_prompt arguments.

Paper Searching

Paper searching is done with the Semantic Scholar API based on your watch terms in watch_terms.txt. Papers are retrieved within the last X days (see --days_back) per watch term. For each paper, the following is done:

• if the paper is flagged as having an available pdf, attempt to retrieve the full pdf
• if pdf retrieval fails, but the paper is still marked as pdf available, try to turn the webpage link into a pdf
• if that also fails, attempt to convert the webpage linked to the paper's DOI into a pdf. (*note: this may sometimes catch a webpage with just an abstract and turn it into a pdf)
• if all else fails, attempt to retrieve the paper abstract and add that as a txt file Note that the default field of study is Computer Science to restrict the paper search, but this can be changed.

Podcast Generation

The script is generated with Gemini in a JSON format. Each "block" of script can be flanked by sound effects provided in the sound effect folder (code and prompt need to be updated to add new sound effects). Script generation doesn't do good when there are too many papers. I'd recommend limiting your papers so you have around 25 at most (and even then it will sort them by interest to avoid generating too long answer). The script JSON can be saved with the --save_script flag.

The audio is generated from the JSON format using the google cloud TTS API. The voice of the podcast can be changed (see the list of voice). When changing the voice (--voice), make sure you also change the language the generation is set to (--language). Intro and Outro are automatically detected to overlay the bgm.mp3 over the text in between those two sound effects (see sound_effects folder). Passing the --no_podcast flag will skip these steps entierly. If you're only interested in paper retrieval and no podcast generation, you can pass this flag (and in the .env file set G_API_KEY to whatever value, it should ignore it but the variable has to exist).

Example Podcast

Here's a podcast generated in May 2025 using the default parameters at the time mp3

python Beholder --help:

usage: Beholder [-h] [-o OUTPUT] [-i INTERESTS [INTERESTS ...]] [-d DAYS_BACK] [-c SEARCH_CHUNK] [-m MODEL] [-t TEMPERATURE] [-s MAX_TOKENS] [-l LANGUAGE] [-v VOICE] [-f FIELDS_OF_STUDY [FIELDS_OF_STUDY ...]]
                [-w WATCH_TERMS] [-p SCRIPT_PROMPT] [--from_papers FROM_PAPERS] [--save_script] [--no_podcast] [--no_audio]

Bot for Extracting & Highlighting Outstanding Literature, Evaluations & Reviews (BEHOLDER).

Requires the following files & directory:
    .env (containing a G_API_KEY variable with a valid google cloud API key for Gemini & TTS)
    watch_terms.txt (a list of terms to search through semantic scholar, one per line)
    script_prompt.txt (the Gemini prompt for script generation)
    sound_effects (a directory containing all available sound effects & bgm)

options:
  -h, --help            show this help message and exit
  -o, --output OUTPUT   folder where papers will be stored and the mp3 will be added
  -i, --interests INTERESTS [INTERESTS ...]
                        areas of interests for the script prompt. Defaults to watch terms
  -d, --days_back DAYS_BACK
                        number of days to look back for papers. Default is 14
  -c, --search_chunk SEARCH_CHUNK
                        paper search limit for each watch term. Higher means more captured results, but can cause issues with Semantic Scholar API. Default is 75
  -m, --model MODEL     gemini model to use. Consider API costs & context windows before changing from default. Default is gemini-1.5-pro
  -t, --temperature TEMPERATURE
                        script generation temperature (higher means more spicy outcomes). Default is 0.7
  -s, --max_tokens MAX_TOKENS
                        maximum token length of generated script (higher means potentially longer scripts, but higher google API cost). Default is 300000
  -l, --language LANGUAGE
                        language of TTS model to use. Need to match the chosen --voice. Default is en-US
  -v, --voice VOICE     voice for podcast. See for a list of available voices (consider API costs per voice engine before making drastic changes). Default is en-US-Chirp3-HD-Algieba
  -f, --fields_of_study FIELDS_OF_STUDY [FIELDS_OF_STUDY ...]
                        fields of study to search in Semantic Scholar. Check API for available fields of study (API paramater fieldsOfStudy). Default is Computer Science
  -w, --watch_terms WATCH_TERMS
                        sets watch_terms.txt path. Defaults to watch_terms.txt
  -p, --script_prompt SCRIPT_PROMPT
                        sets script_prompt.txt path. Defaults to script_prompt.txt
  --from_papers FROM_PAPERS
                        skips paper retrievals, send the given folder of papers to podcast generation
  --save_script         the podcast script's JSON will be saved in the output directory
  --no_podcast          the script will stop after retrieving the papers and will not attempt to generate an audio podcast
  --no_audio            the script will stop after generating the podcast script json. Couple with --save_script to save the script json

Written by Nicolas Jacquin (nicolas.jacquin@umontreal.ca)