Agency research instead of commands

[thejerome1502]

Duplicates

• I have searched the existing issues

Summary 💡

Replace Google Search by an agency (coop of agents) to search both Google and pinecone memory from a prompt sent by autogpt to this agency

Uses multiple instances of a GPT-3.5 model to generate tags for each search engine (e.g., Pinecone and Google) based on the user prompt.
Uses the generated tags to query Pinecone database and Google Search API for the top results relevant to the user prompt.
Uses a GPT-3.5 summarization and organization model to identify the top 5 most relevant search results from Pinecone database and Google Search API, respectively.
Uses a GPT-3.5 summarization and organization model to summarize and organize the search results from Pinecone database and Google Search API in relevance to the user prompt.
Combines the summarized and organized search results from both sources using a GPT-3.5 model.
Outputs it to autogpt

Examples 🌈

No response

Motivation 🔦

No response

[Boostrix]

my 2c: supporting multiple search engines does sound like a good idea (after all, additional search engines would also be a contigency for agents/tasks to try other options), ideally those with an API and a way to return results in a suitable format like JSON/XML - also, being able to search headless without requiring a headless browser would be useful, just the equivalent of curl/wget (wrapped via Python) should suffice ideally

Also, a number of folks mentioned issues relating to browse_website that would not actually conduct a proper search in a looped fashion - i.e. where it would use variations of search term queries and recursively research a topic in order to gather relevant websites and additional keywords. More often than not, "bootstrapping a [re]search" via something like wikipedia is also a good way to get research going. #3635

Realistically, browse_website probably needs to be turned into a plugin and extended - there are a bunch of related RFEs, we probably need some sort of "meta" issue for it:

• Search web only for data post 2021 #1514
• Endless loop of trying to load a bad url. #3444
• Train spawned agents #2944
  Plugin candidate, once the dust settles with Separate the Plugin interface into multiple different interfaces for different types of plugins  #3652

[github-actions]

This issue was closed automatically because it has been stale for 10 days with no activity.