Show HN: R2R V2 – A open source RAG engine with prod features

Hi HN! We're building R2R [https://github.com/SciPhi-AI/R2R https://github.com/SciPhi-AI/R2R], an open source RAG answer engine that is built on top of Postgres+Neo4j. The best way to get started is with the docs - https://r2r-docs.sciphi.ai/introduction https://r2r-docs.sciphi.ai/introduction. This is a major update from our V1 which we have spent the last 3 months intensely building after getting a ton of great feedback from our first Show HN (https://news.ycombinator.com/item?id=39510874 https://news.ycombinator.com/item?id=39510874). We changed our focus to building a RAG engine instead of a framework, because this is what developers asked for the most. To us this distinction meant working on an opinionated system instead of layers of abstractions over providers. We built features for multimodal data ingestion, hybrid search with reranking, advanced RAG techniques (e.g. HyDE), automatic knowledge graph construction alongside the original goal of an observable RAG system built on top of a RESTful API that we shared back in February. What's the problem? Developers are struggling to build accurate, reliable RAG solutions. Popular tools like Langchain are complex and overly abstracted and lack crucial production features such as user/document management, observability, and a default API. There was a big thread about this a few days ago: Why we no longer use LangChain for building our AI agents (https://news.ycombinator.com/item?id=40739982 https://news.ycombinator.com/item?id=40739982) We experienced these challenges firsthand while building a large-scale semantic search engine, having users report numerous hallucinations and inaccuracies. This highlighted that search+RAG is a difficult problem. We're convinced that these missing features, and more, are essential to effectively monitor and improve such systems over time. Teams have been using R2R to develop custom AI agents with their own data, with applications ranging from B2B lead generation to research assistants. Best of all, the developer experience is much improved. For example, we have recently seen multiple teams use R2R to deploy a user-facing RAG engine for their application within a day. By day 2 some of these same teams were using their generated logs to tune the system with advanced features like hybrid search and HyDE. Here are a few examples of how R2R can outperform classic RAG with semantic search only: 1. “What were the UK's top exports in 2023?". R2R with hybrid search can identify documents mentioning "UK exports" and "2023", whereas semantic search finds related concepts like trade balance and economic reports. 2. "List all YC founders that worked at Google and now have an AI startup." Our knowledge graph feature allows R2R to understand relationships between employees and projects, answering a query that would be challenging for simple vector search. The built in observability and customizability of R2R helps you to tune and improve your system long after launching. Our plan is to keep the API ~fixed while we iterate on the internal system logic, making it easier for developers to trust R2R for production from day 1. We are currently working on the following: (1) Improve semantic chunking through third party providers or our own custom LLMs; (2) Training a custom model for knowledge graph triples extraction that will allow KG construction to be 10x more efficient. (This is in private beta, please reach out if interested!); (3) Ability to handle permissions at a more granular level than just a single user; (4) LLM-powered online evaluation of system performance + enhanced analytics and metrics. Getting started is easy. R2R is a lightweight repository that you can install locally with `pip install r2r`, or run with Docker. Check out our quickstart guide: https://r2r-docs.sciphi.ai/quickstart https://r2r-docs.sciphi.ai/quickstart. Lastly, if it interests you, we are also working on a cloud solution at https://sciphi.ai https://sciphi.ai. Thanks a lot for taking the time to read! The feedback from the first ShowHN was invaluable and gave us our direction for the last three months, so we'd love to hear any more comments you have!