Accessing external links

[CybotTM]

Hi,

!bai fasse mir diesen Artikel zusammen: https://members.outcomeedge.com/p/how-to-find-gold-mines

Entschuldigung, ich kann auf externe Links nicht zugreifen. Wenn du möchtest, kannst du mir den Text hier einfügen, und ich helfe dir gerne dabei, eine Zusammenfassung zu erstellen!

Is there any way to baibot to access external links/assets/documents?

[spantaleev]

This sounds like a cool feature, and OpenAI's ChatGPT (their web product) supports such "web browsing". See Should I use ChatGPT or the API?

Unfortunately, it's difficult to emulate. baibot currently only uses the Chat completions API for contacting the model. The model itself cannot do such a thing.

To implement the ability to summarize articles from a question like that:

• the bot would need to understand your intention from your free-form instruction
• it would then need to extract the link
• it would then need to fetch the content at this URL and ideally extract the article text (to avoid feeding ChatGPT tons of useless HTML) - this could be done with Playwright and other tools.. Maybe there are even some APIs that do this "URL to article text" feature
• the bot would then need to feed the article content (and ideally all its images?) to the model via the Chat Completions API and obtain a summary
• if you'd be continuing the discussion in the thread (continuing the discuss the same URL / article text), the bot would also need to persistently store the text somewhere, so it can reuse it. For now the bot only persists state (not too much) via chat messages. Perhaps it would need a proper state database if it's going to persist larger content in a way that's invisible to the user. An alternative would be to extract the article content and post it to the conversation as an attachment file. This way, it can easily reuse it later without going through all steps again.

Given all the uncertainty and complexity, I'd say that this feature is unlikely to get implemented.

[CybotTM]

I know that "chat" is something completely different and more complex than just using the LLM API.
We are also tested our first chat implementation with RAG.

A simple first approach could be extracting the links and attaching the content as context. Or any other way to add context to the chat without pasting wall of text into the chat itself.