RAG (Retrieval-augmented generation) ChatBot

Check out the todo list to see the next steps and improvements we want to implement in this project here.

• Introduction
• Prerequisites
  • Install Poetry
• Bootstrap Environment
  • How to use the make file
  • Environment
• Using the Open-Source LLMs/Embedding Models Locally
  • Supported LLMs Models
  • Supported Embedding Models
• Supported Response Synthesis strategies
• Build the memory index
• Run the Chatbot
• References

This project combines the power of llama.cpp and Chroma 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.

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 Semantic Search models from Sentence Transformers, 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 two 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.
  • 

The Memory Builder builds the vector database in an incremental way, which means that when a document changes, we only update the corresponding chunks in the vector store instead of rebuilding the whole index.

This is achieved through:

• Document-level metadata tracking: every chunk gets tagged with a source doc ID + version hash. When a doc changes, we regenerate chunks for that doc only, delete the old ones by metadata filter, and insert new ones. way cheaper than rebuilding the whole index.
• Incremental ingestion pipeline: the pipeline diffs source docs against what's already indexed (using those version hashes). Only changed/new docs get processed. Keeps compute costs reasonable as the corpus grows.
• Handling deletions: we keep a separate mapping table (doc_id → chunk_ids) in a SQLite db so we can precisely target what to remove without scanning the whole store.

• Python 3.12+
• GPU supporting CUDA 12.4+
• Poetry 2.3.0

For the UI:

• Node 22.12+
• Yarn 1.22+

Install Poetry with pipx 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>

or with pipx:

pipx install poetry==<version> --force

To easily install the dependencies and start the services we created a make file.

• 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.
• Start: make start
  • Start both the backend and frontend ensuring that the backend is running and ready before launching the frontend.
• 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

Copy .𝐞𝐧𝐯.𝐞𝐱𝐚𝐦𝐩𝐥𝐞 → .𝐞𝐧𝐯 and fill it in.

Copy /frontend/.𝐞𝐧𝐯.𝐞𝐱𝐚𝐦𝐩𝐥𝐞 → .𝐞𝐧𝐯 and fill it in.

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.

For the semantic search, we support all the embedding models from Sentence Transformers but we tested those on the table below. To find the list of best embeddings models for the retrieval task in your language (or multiple languages) go to the Massive Text Embedding Benchmark (MTEB) Leaderboard. We do recommend you to use the jina-embeddings-v5-text models, which are small (239M & 677M parameters) with SOTA performance for multilingual retrieval tasks, and they perform very well on the MTEB benchmark.

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

Build the memory index by running:

make migrate_db
python chatbot/memory_builder.py --model-name jinaai/jina-embeddings-v5-text-small-retrieval --chunk-size 1000 --chunk-overlap 50

The Chatbot has a UI built with Vite, React and TypeScript, and a backend built with FastAPI that serves the LLMs through llama-cpp-python.

To install the UI dependencies, run:

cd frontend
nvm use
yarn

# Create .env file
echo "VITE_API_URL=http://localhost:8000" > .env

To start the backend type:

cd backend && PYTHONPATH=.:../chatbot uvicorn main:app --reload

To start the frontend (in a new terminal):

cd frontend && yarn dev

or to start both ensuring that the backend is running and ready before launching the frontend just run:

make start

The application will be available at http://localhost:5173, with the backend API at http://localhost:8000.

You can enable the RAG Mode feature in the UI to ask questions based on the context provided by the Markdown files you loaded and indexed in the previous step:

You can also upload a Markdown file using the file uploader. The document management section shows the uploaded and indexed documents. Once you upload one or multiple files, they will be: uploaded → chunked → embedded → upserted to Chroma.

• Large Language Models (LLMs):
  • LLMs as a repository of vector programs
  • GPT in 60 Lines of NumPy
  • Calculating GPU memory for serving LLMs
  • Introduction to Weight Quantization
  • Uncensor any LLM with abliteration
  • Understanding Multimodal LLMs
  • Direct preference optimization (DPO): Complete overview
• LLM Frameworks:
  • llama.cpp:
    • llama.cpp
    • llama-cpp-python
  • Deepval - A framework for evaluating LLMs:
    • https://github.com/confident-ai/deepeval
  • Structured Outputs
• LLM Datasets:
  • High-quality datasets
• Agents:
  • Agents
  • Building effective agents
• Agent Frameworks:
  • PydanticAI
  • Atomic Agents
    • Want to Build AI Agents? Tired of LangChain, CrewAI, AutoGen & Other AI Agent Frameworks?
  • agno - a lightweight, high-performance library for building Agents.
• 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
  • Qdrant:
    • Qdrant Internals: Immutable Data Structures
    • 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:
  • Open WebUI
    • Running AI Locally Using Ollama on Ubuntu Linux
• Text Processing and Cleaning:
  • clean-text
  • Fast Semantic Text Deduplication
• Inspirational Open Source Repositories:
  • lit-gpt
  • api-for-open-llm
  • AnythingLLM
  • FastServe - Serve Llama-cpp with FastAPI
  • Alpaca
  • LiteLLM