Book and document reading agent

[BradKML]

Continuation of the other point, but more focused on gathering information from then outside (PDFs or Ebooks, Web Articles and Blogs, Video and Audio), caching them, and then writing/modifying thought processes, rather than just experimentation + documentation #136

Some example of multimedia tools (for books, articles, web, video, audio etc.):

• TL;DW Fork as "research multi-tool" https://github.com/rmusser01/tldw
• https://github.com/thiswillbeyourgithub/wdoc
• https://github.com/OE-LUCIFER/Webscout
• https://github.com/SkywalkerDarren/chatWeb

Some example of web crawlers (would be useful as a separate agent):

• Semi-privates https://github.com/mendableai/firecrawl https://github.com/apify/crawlee
• https://github.com/BuilderIO/gpt-crawler
• https://github.com/unclecode/crawl4ai
• https://github.com/adbar/trafilatura
• Chinese Crawler https://github.com/coder-hxl/x-crawl

Thinking some of the major needs:

1. RAGs that has vector storage on different resolution of info (e.g. paragraph, section/article, chapter/category, book/blog)
2. Need for transcription from video and audio (a bit of a stretch but probably is needed)
3. Agents can recursively learn from summarizing with Chain-of-density, and contemplating/documenting
4. Agent has to be able to draw connections between different sources to synthesize new thought

---

Some example Knowedge Chat RAGs (for reference only, not even sure which should be the default):

• https://github.com/open-webui/open-webui
• https://github.com/lobehub/lobe-chat
• Chinese RAGs https://github.com/chatchat-space/Langchain-Chatchat https://github.com/labring/FastGPT
• https://github.com/Mintplex-Labs/anything-llm
• https://github.com/Cinnamon/kotaemon
• https://github.com/arc53/DocsGPT

Some other example local LLMs that includes the occasional text data:

• Top results from Nomic https://github.com/nomic-ai/gpt4all
• https://github.com/mudler/LocalAI
• https://github.com/h2oai/h2ogpt
• Branded as BASB https://github.com/khoj-ai/khoj

[BradKML]

Regarding crawlers there are some other issues:

1. How do you parse webpages and blogs into a more usable format? https://github.com/adbar/trafilatura https://github.com/scrapinghub/article-extraction-benchmark
2. Could fetching archived pages (or archiving pages) from Internet Archive be integrated into this process? e.g. any page that are not archived in 90 days (or ever) will be archived
3. Can crawling be (a) limited by blog domain name, or (b) breath-first and limited to N hops?