lasik-openai-rag ·

Demo LLM (RAG pipeline) web app running locally using docker-compose. LLM and embedding models are consumed as services from OpenAI.

The primary objective is to enable users to ask questions related to LASIK surgery, such as "Is there a contraindication for computer programmers to get LASIK?"

The Retrieval Augmented Generation (RAG) pipeline retrieves the most up-to-date information from the dataset to provide accurate and relevant responses to user queries.

Target setup

The app architecture is presented below:

Sequence diagram:

sequenceDiagram
    User->>Langserve API: query
    Note right of User:  Is there a contraindication <br/>for computer programmers <br/>to get LASIK?
    Langserve API->>OpenAI Embeddings: user query
    OpenAI Embeddings-->>Langserve API: embedding
    Langserve API->>MilvusDB: documents retrieval (vector search)
    MilvusDB-->>Langserve API: relevant documents
    Note right of Langserve API:  Prompt<br/>Engineering...
    Langserve API->>OpenAI LLM: enriched prompt
    OpenAI LLM-->>Langserve API: generated answer

UX:

Prerequisites

• Docker
• An OpenAI key(account should be provisioned with $5, which is the minimum amount allowed)

Quickstart

Build app Docker image:

make app-build

Set your OpenAI API key as environment variable

export OPENAI_API_KEY=<your-api-key>

Spin up Milvus DB:

make db-up

Populate DB with the LASIK eye surgery complications dataset:

make db-populate

Spin-up API:

make app-run

The chatbot is now available at http://localhost:8000/lasik_complications/playground/

Display all available commands with:

make help

Clean up

make clean

Project file structure

├── .github
│   ├── workflow
│   │   └── cicd.yml       <- CI pipeline definition
├── data
│   └── laser_eye_surgery_complications.csv       <- Kaggle dataset
|
├── docs
│   ├── diagrams      <- Folder containing diagram definitions
│   └── img           <-  Folder containing screenshots
│
├── src
│   ├── config.py                  <- Config file with service host/ports or models to be used
│   ├── populate_vector_db.py      <- Scripts that converts texts to embeddings and populates Milvus DB
│   └── server.py                  <- FastAPI/Langserve/Langchain
│
├── .gitignore
├── .pre-commit-config.yaml        <- ruff linter pre-commit hook
├── docker-compose.yml             <- container orchestration
├── Dockerfile                     <- App image definition
├── Makefile                       <- Makefile with commands like `make app-build`
├── poetry.lock                    <- Pinned dependencies
├── pyproject.toml                 <- Dependencies requirements
├── README.md                      <- The top-level README for developers using this project.
└── ruff.toml                      <- Linter config

The dataset

Sourced from Lasik (Laser Eye Surgery) Complications(Kaggle)

Milvus

Milvus is an open-source vector database engine developed by Zilliz, designed to store and manage large-scale vector data, such as embeddings, features, and high-dimensional data. It provides efficient storage, indexing, and retrieval capabilities for vector similarity search tasks.

CICD

• lint: Lints .py files in the repo with ruff
• image-misconfiguration: Detect configuration issues in app Dockerfile (Trivy)
• build: Build app Docker image and push it to the pipeline artifacts
• image-vulnerabilities: App image vulnerablities scanner (Trivy)

Langchain

Langchain is a LLM orchestration tool, it is very useful when you need to build context-aware LLM apps.

Prompt Engineering

In order to provide the context to the LLM, we have to wrap the original question in a prompt template

You can check what prompt the LLM actually received by clicking on "intermediate steps" in the UX

Langserve

LangServe helps developers deploy LangChain runnables and chains as a REST API. This library is integrated with FastAPI.

To do

• The chatbot cannot answer questions related to stats, for example "Are there any recent trends in LASIK surgery complications?", there should be another model that infers the relevant time-window to consider for retrieving the documents and then enrich the final prompt with this time-window.

• Algorithmic feedback with Langsmith. This would allow to test the robustness of the LLM chain in an automated way.

Useful resources

• OpenAI Prompt engineering