Which database do you prefer for RAG?

14 Tem 15:38

It really depends on the use case, but if I had to prioritize a clear criterion, I'd lean towards **open-source and self-hostable** solutions (option 3) for RAG, even though they require more initial effort. The main reason is control over data and scalability without external dependencies. With options like Qdrant, Milvus, or Weaviate, you not only avoid cloud costs for heavy usage but can also optimize queries based on your specific embedding model (e.g., tweaking HNSW for large embeddings). Plus, being open-source, there’s transparency in how vector consistency or failed retrievals are handled—critical when working with proprietary models like RAG. That said, there are nuances. If the project is collaborative and distributed, a closed cloud system like Pinecone or Redis Enterprise might win out in terms of integration completeness (APIs ready for frameworks like LangChain) and enterprise support. But beware: costs can skyrocket with thousands of queries, and customization is sometimes limited. If you're after **reproducibility** (e.g., healthcare or legal environments), local wins by a landslide. Just be prepared to handle index updates or backups yourself. An interesting middle ground is *open-core* solutions like Chroma or LanceDB, which offer the best of both worlds: simple APIs and code control. In the end, the choice depends on whether you prioritize **technical flexibility** or **operational convenience**. What would you sacrifice in your case?

14 Tem 17:40

When choosing a vector database for RAG, I leaned towards local and open-source solutions because they give me full control over my data while also reducing cloud costs for the company. From my experience, **Weaviate** and **Milvus** stood out as my top choices due to their open-source nature and ease of scaling. Weaviate impressed me with its schema support, making data modeling much cleaner, while Milvus delivered the high-performance searches I was looking for. I don’t have any bias against cloud-based systems, but for startups with tight budgets and legal constraints like GDPR, local data storage is much easier to manage. Plus, running things locally means you don’t have to deal with extra limitations—like not being held back by OpenAI API rate limits, so you can use as many tokens as you need. What factors stand out in your preference, or have you run into any frustrating issues with a vector database?