I’ve been poking around the open-source RAG scene and the variety is wild — not just incremental forks, but fundamentally different philosophies.

Quick sketch:

• MiniRAG: ultra-light, pragmatic — built to run cheaply/locally.
• Agent-UniRAG: retrieval + reasoning as one continuous agent pipeline.
• SymbioticRAG: human-in-the-loop + feedback learning; treats users as part of the retrieval model.
• RAGFlow / Verba / LangChain-style stacks: modular toolkits that let you mix & match retrievers, rerankers, and LLMs.

What surprises me is how differently they behave depending on the use case: small internal KBs vs. web-scale corpora, single-turn factual Qs vs. multi-hop reasoning, and latency/infra constraints. Anecdotally I’ve seen MiniRAG beat heavier stacks on latency and robustness for small corpora, while agentic approaches seem stronger on multi-step reasoning — but results vary a lot by dataset and prompt strategy.

There’s a community effort (search for RagView on GitHub or ragview.ai) that aggregates side-by-side comparisons — worth a look if you want apples-to-apples experiments.

So I’m curious from people here who actually run these in research or production:

• Which RAG route gives you the best trade-off between accuracy, speed, and controllability?
• What failure modes surprised you (hallucinations, context loss, latency cliffs)?
• Any practical tips for choosing between a lightweight vs. agentic approach?

Drop your real experiences (not marketing). Concrete numbers, odd bugs, or short config snippets are gold.