RAGCook

This project aims at implementing a simple RAG pipeline based on cooking recipies extracted from marmitton website.

The code uses Qdrant to create a local vectorstore, Langchainand Langgraphlibraries were used to develop the RAG pipeline. To make the whole code easily usable, a Dockerfile is also provided. Finally, the whole code is wrapped into an API using FastAPI.

⚠️ As llama-cpp-python is highly difficult to make work (a lot of building issues) the docker image might not work properly, thus a openai compatibility was added to the project. The steps to use this compatibility is explained later. The current docker will :

• Build llama-cpp from source
• Installs llama-cpp-python

NOTE: These actions will take some time

As the objective was to develop a RAG using local ressources, text pre-processing was not performed using openaiAPIs.

Running the project

To run the project and test it, one should check whether docker is installed, for linux check the following Docker Guide.

Before running the project, for simplicity, git-lfs was not added to the docker image, rather one should download the ggufmodel and place it inside weights/ folder.

NOTE: If you don't download qwen2.5-1.5b-instruct-fp16.gguf consider modifying the config fileconf/llm_global_config.yml folder

Create the docker image then run it

$ docker build -t cookapi .
$ docker run -p 8000:8000 cookapi

Since only the 8000port was exposed you must map this one, if you want to modify the port, you can directly modify the docker image.

Testing the project

import requests
rq  =requests.post(url="http://0.0.0.0:8000/chat", json={"user_message" : "Peut-tu me donner les étapes à suivre si je veux cuisiner des aiguillettes de canard"})
rq.content

Repository architecture

To understand more deeply the project, below the project current architecture

RAGCook/
├── conf
│   ├── llm_global_config.yml
│   └── retrieval_pipeline.yml
├── data
│   ├── RAW_HTML
│   └── VECTORSTORE
├── Dockerfile
├── poetry.lock
├── pyproject.toml
├── README.md
├── src
│   ├── main.py
│   ├── rag_pipeline.py
│   ├── utils
│   │   ├── document_processing.py
│   │   ├── helpers.py  
│   │   └── utils_llm.py
│   └── vectorstore.py
└── weights
    └── DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf

Openai compatbility

To make it work with OpenAI api, one should modify the following file conf/llm_global_config.yml and modify the parameter use_openai_apito True.

Possible Evolutions

A wide range of improvements can be done:

• Implementing test functions
• Largely improve the logger (loguru)
• Implementing evaluation pipeline for both the retrieval part and the generation part
• Transform the Qdrant from local to a service, plus improve the search part. (Adding Hybrid search etc, better indexing)
• Largely improve the RAG workflow to be more efficient and robust (self reflection, Lost in the middle, query rewritting, grading...)
• Try several llm models to asses performances based on the evaluation pipeline