RAG (Retrieval-augmented generation) ChatBot

Important

Disclaimer: The code has been tested on:

• Ubuntu 22.04.2 LTS running on a Lenovo Legion 5 Pro with twenty 12th Gen Intel® Core™ i7-12700H and an NVIDIA GeForce RTX 3060.
• MacOS Sonoma 14.3.1 running on a MacBook Pro M1 (2020).

If you are using another Operating System or different hardware, and you can't load the models, please take a look at the official Llama Cpp Python's GitHub issue.

Warning

• lama_cpp_pyhon doesn't use GPU on M1 if you are running an x86 version of Python. More info here.
• It's important to note that the large language model sometimes generates hallucinations or false information.

Table of contents

• Introduction
• Prerequisites
  • Install Poetry
• Bootstrap Environment
  • How to use the make file
• Using the Open-Source Models Locally
  • Supported Models
• Supported Response Synthesis strategies
• Example Data
• Build the memory index
• Run the Chatbot
• Run the RAG Chatbot
• How to debug the Streamlit app on Pycharm
• References

Introduction

This project combines the power of Lama.cpp, Chroma and Streamlit to build:

• a Conversation-aware Chatbot (ChatGPT like experience).
• a RAG (Retrieval-augmented generation) ChatBot.

The RAG Chatbot works by taking a collection of Markdown files as input and, when asked a question, provides the corresponding answer based on the context provided by those files.

Note

We decided to grab and refactor the RecursiveCharacterTextSplitter class from LangChain to effectively chunk Markdown files without adding LangChain as a dependency.

The Memory Builder component of the project loads Markdown pages from the docs folder. It then divides these pages into smaller sections, calculates the embeddings (a numerical representation) of these sections with the all-MiniLM-L6-v2 sentence-transformer, and saves them in an embedding database called Chroma for later use.

When a user asks a question, the RAG ChatBot retrieves the most relevant sections from the Embedding database. Since the original question can't be always optimal to retrieve for the LLM, we first prompt an LLM to rewrite the question, then conduct retrieval-augmented reading. The most relevant sections are then used as context to generate the final answer using a local language model (LLM). Additionally, the chatbot is designed to remember previous interactions. It saves the chat history and considers the relevant context from previous conversations to provide more accurate answers.

To deal with context overflows, we implemented three approaches:

• Create And Refine the Context: synthesize a responses sequentially through all retrieved contents.
  • 
• Hierarchical Summarization of Context: generate an answer for each relevant section independently, and then hierarchically combine the answers.
  • 
• Async Hierarchical Summarization of Context: parallelized version of the Hierarchical Summarization of Context which lead to big speedups in response synthesis.

Prerequisites

• Python 3.10+
• GPU supporting CUDA 12.1+
• Poetry 1.7.0

Install Poetry

Install Poetry with the official installer by following this link.

You must use the current adopted version of Poetry defined here.

If you have poetry already installed and is not the right version, you can downgrade (or upgrade) poetry through:

poetry self update <version>

Bootstrap Environment

To easily install the dependencies we created a make file.

How to use the make file

Important

Run Setup as your init command (or after Clean).

• Check: make check
  • Use it to check that which pip3 and which python3 points to the right path.
• Setup:
  • Setup with NVIDIA CUDA acceleration: make setup_cuda
    • Creates an environment and installs all dependencies with NVIDIA CUDA acceleration.
  • Setup with Metal GPU acceleration: make setup_metal
    • Creates an environment and installs all dependencies with Metal GPU acceleration for macOS system only.
• Update: make update
  • Update an environment and installs all updated dependencies.
• Tidy up the code: make tidy
  • Run Ruff check and format.
• Clean: make clean
  • Removes the environment and all cached files.
• Test: make test
  • Runs all tests.
  • Using pytest

Using the Open-Source Models Locally

We utilize the open-source library llama-cpp-python, a binding for llama-cpp, allowing us to utilize it within a Python environment. llama-cpp serves as a C++ backend designed to work efficiently with transformer-based models. Running the LLMs architecture on a local PC is impossible due to the large (~7 billion) number of parameters. This library enable us to run them either on a CPU or GPU. Additionally, we use the Quantization and 4-bit precision to reduce number of bits required to represent the numbers. The quantized models are stored in GGML/GGUF format.

Supported Models

🤖 Model	Supported	Model Size	Max Context Window	Notes and link to the model card
llama-3.2 Meta Llama 3.2 Instruct	✅	3B	128k	Recommended model optimized to run locally on a mobile or edge device - Card
llama-3.1 Meta Llama 3.1 Instruct	✅	8B	128k	Recommended model Card
openchat-3.6 - OpenChat 3.6	✅	8B	8192	Card
openchat-3.5 - OpenChat 3.5	✅	7B	8192	Card
starling Starling Beta	✅	7B	8192	Is trained from Openchat-3.5-0106. It's recommended if you prefer more verbosity over OpenChat - Card
phi-3.5 Phi-3.5 Mini Instruct	✅	3.8B	128k	Card
stablelm-zephyr StableLM Zephyr OpenOrca	✅	3B	4096	Card

Supported Response Synthesis strategies

✨ Response Synthesis strategy	Supported	Notes
create-and-refine Create and Refine	✅
tree-summarization Tree Summarization	✅
async-tree-summarization - Recommended - Async Tree Summarization	✅

Example Data

You could download some Markdown pages from the Blendle Employee Handbook and put them under docs.

Build the memory index

Run:

python chatbot/memory_builder.py --chunk-size 1000 --chunk-overlap 50

Run the Chatbot

To interact with a GUI type:

streamlit run chatbot/chatbot_app.py -- --model llama-3.2 --max-new-tokens 1024

Run the RAG Chatbot

To interact with a GUI type:

streamlit run chatbot/rag_chatbot_app.py -- --model llama-3.2 --k 2 --synthesis-strategy async-tree-summarization

How to debug the Streamlit app on Pycharm

References

• Large Language Models (LLMs):
  • Calculating GPU memory for serving LLMs
  • GPT in 60 Lines of NumPy
  • Introduction to Weight Quantization
  • Uncensor any LLM with abliteration
• LLM Frameworks:
  • llama.cpp:
    • llama.cpp
    • llama-cpp-python
  • Ollama:
    • Ollama
    • Ollama Python Library
    • Analysis of Ollama Architecture and Conversation Processing Flow for AI LLM Tool
    • How to Customize Ollama’s Storage Directory
• Agent Frameworks:
  • PydanticAI
• Embeddings:
  • all-MiniLM-L6-v2
    • This is a sentence-transformers model: It maps sentences & paragraphs to a 384 dimensional dense vector space and can be used for tasks like clustering or semantic search.
• Vector Databases:
  • Indexing algorithms:
    • There are many algorithms for building indexes to optimize vector search. Most vector databases implement Hierarchical Navigable Small World (HNSW) and/or Inverted File Index (IVF). Here are some great articles explaining them, and the trade-off between speed, memory and quality:
      • Nearest Neighbor Indexes for Similarity Search
      • Hierarchical Navigable Small World (HNSW)
      • From NVIDIA - Accelerating Vector Search: Using GPU-Powered Indexes with RAPIDS RAFT
      • From NVIDIA - Accelerating Vector Search: Fine-Tuning GPU Index Algorithms

      • PS: Flat indexes (i.e. no optimisation) can be used to maintain 100% recall and precision, at the expense of speed.

  • Chroma
    • chroma
  • Food Discovery with Qdrant
• Retrieval Augmented Generation (RAG):
  • Building A Generative AI Platform
  • Rewrite-Retrieve-Read

    • Because the original query can not be always optimal to retrieve for the LLM, especially in the real world, we first prompt an LLM to rewrite the queries, then conduct retrieval-augmented reading.

  • Rerank
  • Building Response Synthesis from Scratch
  • Conversational awareness
  • RAG is Dead, Again?
• Chatbot UI:
  • Streamlit:
    • Build a basic LLM chat app
    • Layouts and Containers
    • st.chat_message
    • Add statefulness to apps
      • Why session state is not persisting between refresh?
    • st.cache_resource
    • Handling External Command Line Arguments
  • Open WebUI
    • Running AI Locally Using Ollama on Ubuntu Linux
• Text Processing and Cleaning:
  • clean-text
• Inspirational Open Source Repositories:
  • lit-gpt
  • api-for-open-llm
  • AnythingLLM
  • FastServe - Serve Llama-cpp with FastAPI