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.

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's GitHub issue.

Warning

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
  • Environment
  • Set the Open-Source LLM Model
  • Set the Embedding Model
  • Set the Response Synthesis strategy
• Build the memory index
• Run the Chatbot
• References

Introduction

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.

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 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.

Important

One thing to watch out for — if you ever swap embedding models, you must rebuild it from scratch since the vector spaces won’t be compatible. Plan for that early.

Prerequisites

• Python 3.12+
• GPU supporting CUDA 12.4+ or Apple Silicon M-series
• Poetry 2.3.0+
  • See notes/poetry.md.
• Docker 24.0.6+ and Docker Compose 5.0.2+
• NVIDIA Container Toolkit installed (optional, for CUDA support)
  • See notes/llama-server-docker.md.

For the UI:

• Node 22.12+
• Yarn 1.22+

Bootstrap Environment

To easily install the dependencies and start the services 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.
  • Both starts llama.cpp server locally via Docker compose.
• Start: make start
  • Start both the backend and frontend ensuring that the backend is running and ready before launching the frontend.
• Start llama.cpp server
  • on CUDA: make start_llama_server_cuda
  • on Metal: make start_llama_server_metal
  • Start the llama.cpp server locally via Docker compose.
  • It will be available at http://0.0.0.0:8080 (it will show the llama-ui).
• Stop llama.cpp Server: make stop_llama_server
  • Stop the llama.cpp server if it's running locally.
• 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

Environment

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

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

To install the UI dependencies, run:

cd frontend
nvm use
npm install -g yarn
yarn

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

Set the Open-Source LLM Model

llama-cpp serves as a C++ backend designed to work efficiently with transformer-based models, which runs either on a CPU or GPU. It uses quantized models which are stored in GGML/GGUF format.

We can load whatever GGUF model we want from HuggingFace.

In the .𝐞𝐧𝐯 we need to set the MODEL variable with the name of the model we want to load, and the MODEL_URL variable with the URL of the model in GGUF format:

MODEL="Meta-Llama-3.1-8B-Instruct-Q4_K_M"
MODEL_URL="https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf"

Important

The Chatbot must be restarted after changing the model.

The chosen model will be downloaded in the /models folder and loaded in the llama.cpp server.

Note

We can load models that fits our hardware capacity and speed requirements. To decide which hardware to use/buy to host local LLMs we recommend to read this great benchmarks:

• Performance of llama.cpp on Nvidia CUDA
• Performance of llama.cpp on Apple Silicon M-series

Decision model:

• Memory capacity is the main limit. Check if the model fits in memory (with quantization):
  • CanIRun.ai - Find out which AI models your machine can actually run.
  • whichllm - Auto-detects your GPU/CPU/RAM and ranks the top models from HuggingFace that fit your system.
• Memory bandwidth mostly determines speed (tokens/sec). Check if the bandwidth gives an acceptable speed.
• If not, upgrade hardware or optimize the model.

For instance, it seems better to buy a second-hand or refurbished Mac Studio M2 Max with at least 64GB RAM, since it has 400Gbps of memory bandwidth compared to the M4 Pro, which has just 273Gbps.

We recommend to start with Qwen 3.5 9B or Meta Llama 3.2 Instruct 3B since they are small enough to run on a cheap GPU with 6GB of VRAM and try larger models like gpt-oss 120B if you have the right capacity.

We also recommend few models to start in the table below.

🤖 Model	Model Size	Max Context Window	Notes and link to the model card
Qwen 3.6 27B	27B	262k	Recommended model Card
Qwen 3.6 35B A3B	35B (3B activated)	262k	Card
Qwen 3.5 0.8B	0.8B	256k	Tiny and fast multimodal, great for edge device - Card
Qwen 3.5 2B	2B	256k	Multimodal for lightweight agents (small tool calls) - Card
Qwen 3.5 4B	4B	256k	Doesn’t drift from tasks as bad as 2B Card
Qwen 3.5 9B	9B	256k	Recommended model Can handle more complex tasks and competes with larger models like gpt-oss 120B Card
Meta Llama 3.2 Instruct	1B	128k	Optimized to run locally on a mobile or edge device - Card
Meta Llama 3.2 Instruct	3B	128k	Optimized to run locally on a mobile or edge device - Card
Meta Llama 3.1 Instruct	8B	128k	Old but still recommended Card
DeepSeek R1 Distill Qwen 7B	7B	128k	Experimental Card

Set the Embedding Model

For the semantic search, we support all the embedding models from Sentence Transformers but we tested those on the table below.

In the .𝐞𝐧𝐯 we need to set the EMBEDDING_MODEL variable with the name of the model we want to load:

EMBEDDING_MODEL="all-MiniLM-L6-v2"

To find the list of best embeddings models for the retrieval task in the language (or multiple languages) go to the Massive Text Embedding Benchmark (MTEB) Leaderboard. We do recommend 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.

🧠 Embedding Model	Supported	Model Size	Max Tokens	Retrieval score (MTEB)	Notes and link to the model card
all-MiniLM-L6-v2 - Sentence Transformers All MiniLM L6 v2	✅	0.023B	512	33.30	Card
all-MiniLM-L12-v2 - Sentence Transformers All MiniLM L12 v2	✅	0.033B	256	33.37	Card
all-mpnet-base-v2 - Sentence Transformers All Mpnet base v2	✅	0.109B	384	33.80	Card
jinaai/jina-embeddings-v5-text-small-retrieval - jina-embeddings-v5-text-small	✅	0.596B	32k	64.88	Recommended model Card
jinaai/jina-embeddings-v5-text-nano-retrieval - jina-embeddings-v5-text-nano	✅	0.212B	8k	63.26	Card

Set the Response Synthesis strategy

In the .𝐞𝐧𝐯 we need to set the SYNTHESIS_STRATEGY variable with the name of the strategy we want to use for the response synthesis:

SYNTHESIS_STRATEGY="tree-summarization"

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

Build the memory index

We can download some Markdown pages from the Blendle Employee Handbook and put them under docs.

Build the memory index by running:

make migrate_db
cd scripts && PYTHONPATH=.:../backend python memory_builder.py --model-name jinaai/jina-embeddings-v5-text-small-retrieval --chunk-size 1000 --chunk-overlap 50

Run the Chatbot

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

To start both the backend and frontend 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.

We 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:

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

References

• 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:
  • 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?
• Text Processing and Cleaning:
  • clean-text
  • Fast Semantic Text Deduplication
• Inspirational Open Source Repositories:
  • lit-gpt
  • api-for-open-llm
  • AnythingLLM
  • Alpaca
  • LiteLLM