PureCPP

This is an open-source modular RAG (Retrieval-Augmented Generation) system developed by PureAI, to which I contributed during my internship, integrating C++ and Python, deploying to PyPI, and optimizing for high-performance vector processing and querying. It also includes careful memory control strategies to maximize performance, culminating in the development of a functional and scalable vector database engine. In addition to its modular architecture, the project features orchestration and optimization through build and deploy pipeline scripts, reducing processes that originally took ~3 hours down to just 30 minutes.

Note

As this project is extensive and contains many modular components, this documentation will initially focus on explaining the parts I developed in the ChunkDefault and ChunkQuery modules, as well as their integration with the VectorDatabase.

Later sections will extend the documentation to cover:

• The C++ <-> Python bindings 🔗Purecpp Refactor: Scratch Pad ;
• The modular CMake architecture🔗Purecpp Refactor: Scratch Pad ;
• 🔗 Automated PIP build and legacy deployment system.

---

Table of Contents

• Introduction
• Main Contributions
• Related Files and Areas
• Technologies Used
• Functionality
• Disclaimer

Main Contributions

• Implemented nearly all C++ <-> Python bindings, enabling seamless integration between the core C++ and Python.
• Implemented and optimized the build-and-deploy pipeline, reducing process time from ~3 hours to just 30 minutes.
• Created a Docker environment for isolated, reproducible development and usage.
• Restructured the repository for better modularity, scalability, and maintainability.
• Designed and implemented the core vector database engine, ensuring scalable, functional vector storage and retrieval.
• Developed cosine similarity query implementations, integrated with FAISS for optimized similarity search.
• Developed of the backbone architecture of the core query system, focusing on efficient construction and high-performance execution.

Related Files and Areas

The Preview:🔗Purecpp Refactor: Scratch Pad , under tests before its official release under the PURE Ecosystem.

• All PIP_* folders:

  • Scripts and configurations for Python packaging and deployment.

• All automation scripts (*.sh):

  • Build, deploy, and environment orchestration scripts.

• Core modules:

  • VectorDatabase
  • ChunkDefault
  • ChunkQuery

• Documentation:

  • All documentation related to my file/directories

• Bindings:

  • All files related to C++/Python bindings integration

• CMake:

  • All files and configurations with uppercase CMake in their name, including modular build configurations

Technologies Used

• CMake — Modular build system configuration.
• Conan — C++ package and dependency management.
• Torch / PyTorch — Machine learning and tensor operations.
• C++ — Core implementation language.
• Rust — Auxiliary or experimental components.
• Docker — Containerization for reproducible environments.
• ManyLinux — Multi-Python-version builds and compatibility for pip wheels.
• Ninja — Fast build system used with CMake for compilation.
• PyBind11 — C++ <-> Python bindings.
• FAISS — Vector similarity search (used in cosine similarity, L2, inner product queries).
• RE2 — High-performance regular expressions.
• OpenAI API (EmbeddingOpenAI) — For generating embeddings via models like text-embedding-ada-002.
• OpenMP — Parallelization for multi-threaded C++ code.
• OpenBLAS — Optimized linear algebra backend for matrix operations.
• ONNX — Open Neural Network Exchange for model interoperability.

FUNCTIONALITY

Core Vector Database Engine

I engineered and integrated the fundamental data structures that gave rise to the engine powering the Vector Database — scalable and high-performance core.

Focusing on:

• ChunkDefault
• ChunkQuery
• VectorDatabase

High-Level Functionality Demonstration

I have created a Jupyter Notebook that explains the high-level functionality of this components that I implemented This notebook demonstrates how these components work together and highlights the key aspects of the three main libraries involved. or Click here to access the notebook directory

Low-Level Functionality

NOTE:
Initially, I created these structures and functions in the directory Chunk/ChunkCommons/ChunkCommons.h.
I will provide further details and explanations about this part later.

---

---

---

Disclaimer

These contributions reflect technical solutions I personally implemented, outside the formal scope of my internship contract, using no internal company resources. No confidential or proprietary information is disclosed here; all content is based on public open-source work.

This repository contains code I developed during my time collaborating with PureAI. Officially, my role was focused on testing tasks; however, over time, I contributed far beyond what was initially requested, because is a startup.

I developed:

• the backbone architecture for VDB engine,
• C++ <-> Python bindings,
• modular chunk and embedding systems,
• all pypi integration
• and automation scripts for build and deployment.

While formally an intern, I undertook responsibilities typically associated with intermediate-level engineering roles.

License & Visibility

The repository is public under the MIT license. Everything shared here reflects parts I personally worked on or played a significant role in developing. There is no confidential or proprietary information being revealed here:

1. The code is already publicly available,
2. My intent is to explain and clarify the parts I directly contributed to,
3. I aim to provide technical transparency and highlight the complexities often hidden behind raw code.