AI Dev Team

[ElishaKay]

Setup:

Step 1: Generate a Github Personal Access Token

For the GITHUB_TOKEN (Personal Access Token) value, you'll need to generate a Personal Access Token:
Click on your profile picture in the top-right corner of GitHub and select "Settings".
In the left sidebar, click on "Developer settings".
Click on "Personal access tokens", then "Tokens (classic)".
Click "Generate new token" and select the appropriate scopes for your needs.
Copy the generated token.
Paste the token you generated into the value field of the new secret.
Click "Add secret" to save it.

Step 2: Set these environment variables:

OPENAI_API_KEY={Your OpenAI API Key here}
TAVILY_API_KEY={Your Tavily API Key here}

PGVECTOR_CONNECTION_STRING=postgresql://username:password...
GITHUB_TOKEN=

Step 3:

pip install -r multi_agents/requirements.txt
python -m multi_agents.dev_team.main

Status: Running to completion

[assafelovic]

This is sick @ElishaKay have to admit I'm hooked on this!

[assafelovic]

Just please try to reuse any already existing code i see a lot of duplicates!

[ElishaKay]

Status report:
The bot supports 1-on-1 chats.
Within discord, the user can trigger a modal by typing: "/ask" in the chat and hitting "Enter"
Here is what the form looks like

[ElishaKay]

Status Report:

Flow:

Step 1: The GithubAgent is in charge of the first step of fetching the data from Github with an API Tool & logging the Directory Structure.

He's also in charge of saving the Github repo within a LangChain VectorStore.

Step 2: The RepoAnalyzerAgent leverages the vectorstore with the repo by running GPTResearcher like so:

   # Run GPTResearcher
   researcher = GPTResearcher(
       query=query,
       report_type="research_report",
       report_source="langchain_vectorstore",
       vector_store=vector_store,
   )

Step 3: The WebSearchAgent takes the output of the RepoAnalyyzerAgent & complements any insights from his analysis with info from the web. He runs GPTR like so:

   # Run GPTResearcher
   researcher = GPTResearcher(
       query=query,
       report_type="research_report",
       report_source="web"
   )

Step 4: The RubberDuckerAgent talks out loud about what the game plan is for answering the user. He's forced to talk through his reasoning based on the outputs of the RepoAnalyyzerAgent & WebSearchAgent

The GithubAgent now has the ability to parse & embed the entire Github Repo contents in optimized Langchain Documents format. He passes those Langchain Documents to the VectorAgent who saves them in the LangChain VectorStore.

That VectorStore is then passed into the GPTR report flow by the RepoAnalyzerAgent, like so:

researcher = GPTResearcher(
      query=query,
      report_type="research_report",
      report_source="langchain_vectorstore",
      vector_store=vector_store,
)

The resulting report is solid, but:

• properly embedding the full github repo takes a minute or so

Areas for improvement:

• leverage the same vector_store across reports & follow-up questions (i.e. don't re-embed the entire repo on every run)
• add the WebSearch Agent into the mix
• Save the file names from the results of the vector search within Langgraph State (to be leveraged by FileSearchAgent and/or RubberDuckerAgent)
• Give the Github Agent a method to calculate the delta between commits & only re-embed the files that changed since the latest commits on a given branch

Also:

• rebased on top of Add Document To Vector Store #838
• would be cool to also rebase this on top of the follow-up questions feature in [Experimental] Chat with history #832 before the big release

[ElishaKay]

Current flow

[assafelovic]

@ElishaKay whats up brother, what's the status on this? Please also see we've massively refactored the code base so there are some conflicts here

[ElishaKay]

@assafelovic

Status Update:

Langchain PGVector support is looking good for the saving & retrieving parts of the flow.

That will be an important step for long-form documents that don't need to re-embed on every run & seamless support for langgraph cloud.

Next steps are what's marked here in "Areas for improvement". I'll probably end up resolving the merge conflicts after those are implemented.