🔎 Multi Agentic Research (RAG)

• Multi-agentic researcher researches the query/topic from documents fetched from Tavily Search Engine.
• The output is the report about the query/topic based on the retrieved documents.
• The graph ends when the hallucination point is 1, or else the graph restarts from collecting documents from Tavily for generated queries to retrieve related info.

Demo

• To improve performance and avoid reaching limit of free tier of Google Generative AI, I limited the length of the response/report to be min 200 words.

Work in progress

1. INTERNAL DOCUMENTATION
2. Access to documents to reliable sources for academic sources (sciencedirect, elsevier, springer, ...)
3. Construct more detail report (e.g Write report for each query)
4. Human input (e.g research plan, sources to fetch documents for queries)

Installation

Customize LLMs, retrievers and report length in config.yaml

1. Install Python 3.12 or later. Guide.

2. Clone the project and navigate to the directory:

   git clone https://github.com/maihoangbichtram/multi_agentic_rag
   cd multi_agentic_rag

3. Set up API keys by exporting them or storing them in a .env file.

   Copy .env.example to .env

4. Create and Activate virtual environment

   4.1 In terminal: Create a virtual environment

   python3 -m venv .venv

   Activate the virtual environment

   . .venv/bin/activate

   4.2 Poetry

   poetry shell

5. Install dependencies and start the server:

   5.1 In terminal:

   pip install -r requirements.txt

   5.2 Poetry

   poetry install

6. Start the application

   6.1 In terminal:

   python app/main.py

   6.2 Poetry

   poetry run uvicorn app.main:app --reload --timeout-keep-alive 720

Deployment

Google Cloud

Prerequisites

• Google Cloud Account
• Google CLI
• Service account under IAM (to access to Google Cloud services.)
• (Access) Key downloaded as json (for the service account)

Terraforming GC Components

Tf files

• main.tf
• variables.tf
• terraform.tfvars (not public): Declare values for variables

  #  GCP settings
  project_id = "project_id"
  
  # GCP region
  region = "europe-west3"
  
  #  Artifact registry repository
  registry_id = "registry_id"

Commands

Run commands in order

• terraform init
• terraform plan
• terraform apply: create the GCP resources

Note

Enable GCP APIs manually if Terraform reports errors with permission:

gcloud services enable artifactregistry.googleapis.com
gcloud services enable run.googleapis.com

Deploying to cloud

• Create requirements.txt for docker file

poetry export -f requirements.txt --output requirements.txt

Build and test docker image locally

docker build --pull --rm -f "Dockerfile" -t <image_name>:latest "."

docker run --rm -it -p 8080:8080/tcp <image_name>:latest

Build docker image for Cloud Run

• Create builder

docker buildx create --name <builder_name> --bootstrap --use

• Build the image

docker buildx build --file Dockerfile \
  --platform linux/amd64 \
  --builder <builder_name> \
  --progress plain \
  --build-arg DOCKER_REPO=<gcd_region>-docker.pkg.dev/<gcd_project_id>/<gcd_repository_name>/ \
  --pull --push \
  --tag <gcd_region>-docker.pkg.dev/<gcd_project_id>/<gcd_repository_name>/<image_name>:latest .

Deploy

Added necessary API keys for the app to run probably. If you configure OpenAI for llm, add env var OPENAI_API_KEY.

gcloud run deploy app \                             
--image <gcd_region>-docker.pkg.dev/<gcd_project_id>/<gcd_repository_name>/<image_name>:latest \
--region <gcd_region> \
--platform managed \
--allow-unauthenticated \
--set-env-vars GOOGLE_API_KEY=<GOOGLE_API_KEY>,TAVILY_API_KEY=<TAVILY_API_KEY>,CO_API_KEY=<CO_API_KEY>