Introduction

Chatbot option:

A customizable GenAI RAG application using Google Cloud components:

Gemini: LLM
textembedding-gecko: embedding model
Cloud Run: app deployment
DocAI: document parsing

The app also uses the following open source components:

gradio: UI
langchain: orchestration of RAG components

The project has quick customizable options in src/config.py:

unstructured text parsing
table parsing
entity parsing
advanced options for each parsing strategy
chatbot vs search function

Roadmap:

✅ Add retrieval snippets to UI
✅ Integrate DocAI Form Parser for tables. Remove table snippets from unstructured text parsing
✅ Integrate DocAI GenAI CDE 2.0 parser for entity extraction
✅ Add option for chatbot instead of single search
⬜ Read directory instead of one file
⬜ Read files from GCS
⬜ Include Vertex Search as a retrieval source
⬜ Include Vertex Vector Search as a retrieval source
⬜ Filtered retrieval based on document metadata (year, customer, etc...)
⬜ Evaluation job of RAG model for accuracy
⬜ Add github actions

Requirements

Google Cloud Platform (GCP) project python version >= 3.8

Setup

Google Cloud Project Setup

Enable the following APIs

Cloud Document AI API
Vertex AI API
Cloud Run API
Generative Language API

Create Service Account with following roles:

Vertex AI Administrator
Document AI API User
Cloud Run Developer

Create a service account key which automatically downloads a json file with your key:

Next, edit src/config.py to include your GCP project:

PROJECT = <PROJECT_ID>

Setup Documents and Parser

Add the pdf document that you are trying to ask questions from to the "docs/" directory and modify src/config.py:

FILENAME=<DOCUMENT_LOCATION>

Create a DocAI processor and modify src/config.py:

For this project you will need two processors of two different types: One is the FORM_PARSER_PROCESSOR (setup) and the CUSTOM_EXTRACTION_PROCESSOR (setup). The form parser will be used to parse unstructured text blocks and tables while the custom extraction processor will be used to pull entities. Note that the custom extraction processor has some new Generative AI features that make it easier to extract entities.

FORM_PROCESSOR_ID=<ID> 
FORM_PROCESSOR_VERSION=<VERSION>
CDE_PROCESSOR_ID=<ID>
CDE_PROCESSOR_VERSION=<VERSION>

Running the App Locally or in Cloud Run

Option: Runing Locally

Find the location of the json service account file that was downloaded and run on terminal:

export GOOGLE_APPLICATION_CREDENTIALS = <SERVICE ACCOUNT FILE LOCATION>

Clone this terminal and run:

cd src
python main.py

Option: Running in Cloud Run

Modify Dockerfile to include your own project:

ENV GOOGLE_CLOUD_PROJECT <PROJECT_ID>

Deploy to cloud run:

gcloud run deploy <ANY APP NAME> --source . --region <REGION> --service-account=<SERVICE_ACCOUNT>

felipecastrillon/GoogleCloudGenAgent