Feat: Enhance question-answering context by processing search result web pages

[Swiftyos]

Is your feature request related to a problem? Please describe.
Currently, the question-answering system relies on the context provided by the search results alone. This can sometimes lead to incomplete or less accurate answers, as the context may not be comprehensive enough.

Describe the solution you'd like
Implement a feature that allows the system to open the web pages from the search results and process their content to enhance the context used for answering questions. This would involve:

1. Retrieving the web pages from the search results
2. Extracting relevant text and information from the web pages
3. Incorporating the extracted information into the existing context
4. Using the enhanced context to generate more accurate and comprehensive answers

Describe alternatives you've considered

1. Relying on search result snippets: This alternative would involve using only the snippets provided by the search engine results. However, this may not provide enough context for complex questions.
2. Using browserless.io: This is a service that allows for web page rendering and interaction, but it can be expensive for large-scale use.
3. Employing Selenium: Selenium is a powerful tool for web automation and scraping, but it is quite heavy-weight and may not be the most efficient solution for this use case.

Additional context
Implementing this feature would greatly improve the question-answering capabilities of the system by providing a richer context derived from the actual content of the web pages.

[Swiftyos]

I don't feel there is a clear solution on how to do this well. Would be great to hear from other people about what has and hasn't worked for them.

[ItzCrazyKns]

Hi, the co-pilot feature I'm currently developing processes multiple queries to gather extensive information on a topic from the web. We use SearxNG, a meta-search engine, which returns snippets based on keyword searches. For all processed queries, we likely have comprehensive data on the topic. This data is further cleaned and reranked using similarity searches and certain thresholds.

If you're specifically referring to scraping entire web pages, I've already explored this approach. Although I managed to achieve the lowest possible latency, the process of performing multiple queries, visiting pages, extracting information, and reranking it was not only time-consuming but also computationally expensive.

I'll convert this into a discussion rather than an issue for now.

[Wadera]

I would love to have it as an option. Even being computing heavy and time consuming - it might significantly increase research value, as searxng snippets are not telling full story and don't allow to respond correctly to some complex searching. And even waiting 3-5 mins to process multiple pages is still significantly quicker than manually searching trough websites manually or manually copy-paste long articles into chat to feed it with right context. Imagine you are writing an ap and instead of feeding AI chat with service documentation you can simple ask a question and peplexica would "scrape" a full documentation of the project and then allow to design correct API call for your code. That would be brilliant and I could wait a few mins while it does this job for me ;)

[Swiftyos]

If you're able to solve this problem, it would put your project at the centre of open source web agents for sure. It's a problem we've been grappling with for AutoGPT and do have a solution, just wondering if a project focused specifically on web retrieval will be able to solve it even better.

SearxNG is pretty fantastic! ❤️