Building a better RAG pipeline: Introducing the open-source RAG-Framework-2026 💡

Gemini API
, ,
1

Hi everyone,

I’m sure many of us in the community are using LLMs (especially Gemini) to tackle RAG (Retrieval-Augmented Generation) challenges. However, transitioning from a Proof of Concept (PoC) to a production-ready application often brings up hurdles related to code structure, context management, and optimizing the data ingestion pipeline.

To address these pain points and save fellow developers some time, I want to share a project I’ve been working on: RAG Framework 2026.

:bullseye: Key Highlights of the Framework:

  • Modular Design: Easily swap out or upgrade components (Embedding models, Vector DBs, LLMs) without breaking the rest of your system.

  • Optimized for Google AI: Designed to be plug-and-play with Google APIs, helping you easily maximize Gemini’s potential.

  • Clear Processing Pipelines: Smoothly handles the entire flow from document ingestion to user querying.

:fire: My Goal: I hope this repo serves as a valuable resource—making it easier for newcomers to approach RAG, while saving experienced developers hours of boilerplate setup.

:link: You can check out the code and clone it here: https://github.com/Taitv01/rag-framework-2026.git

Feel free to fork it and play around! If you find it useful, dropping a :star: on the repo would be a huge motivation boost. I also highly welcome Pull Requests or Issues from the community so we can make this framework even better together.

Thanks for reading!

2

Hi team, the system has been fully updated with all items from Phase A to F. Here are the latest changes and new features:

:glowing_star: Highlights

  • :brain: adaptive_rag.py (776 lines): Automatically selects the optimal pipeline based on the complexity of the question.

  • :star: Default Embedding Updated: Switched to BAAI/bge-m3 – the best baseline for Vietnamese RAG in 2026.

  • :deciduous_tree: Advanced Chunking Techniques: Integrated state-of-the-art techniques including RAPTOR and Late Chunking.

:white_check_mark: Completed Items Summary

  • Bug Fixes & Code Optimization: Resolved 3 critical bugs (missing imports, typos) and 6 code quality issues (CORS, duplication, side effects…).

  • Model Catalog: Updated the catalog (set BGE-M3 as default, added 3 new embeddings and 2 rerankers).

  • Documentation: Fully updated PHASES.md and README.md.

:light_bulb: Important Note: There are 5 new dependencies added in version 1.1.0. After pulling the latest code, please make sure to run the following command to install them:

pip install -r requirements.txt