META’s New SIRA: Superintelligence RAG

Imagine you have lost your keys in a dark, unfamiliar room. How do you find them? You don't sit in the hallway and try to hallucinate the exact atomic structure of the room to guess where they are. Nor do you simply run around the room at random, hoping you eventually step on them and get a "reward." Yet, this is exactly what the two reigning titans of AI search are currently doing. In one corner, we have SIRA (Superintelligent Retrieval Agent). It attempts to sit in the hallway and sketch the room. In the other corner, we have Search-R1. It runs around the room, learning not what the room looks like, but how to swing its arms so it doesn't hit a wall.

This is the right way.. just need the mechanisim to search the db prior to rag search and retrieve things..

"Superintelligent Retrieval Agent" = BM25 with cached query expansions. The branding does all the heavy lifting.

"Results are impressive but on standard benchmarks — real-world noisy corpora, adversarial queries, or massive scale may differ. "Superintelligent" is marketing flair; it's a smart, expert-mimicking pipeline, not AGI-level retrieval." - Grok

Since when does 69% average success => "Superinteligence"?

Wow. It just gets worse the more you look at it

It's paradox, why superintelligence needs RAG for knowledge confirmation? no different with previous concepts

I once had the idea that AI systems could analyze audio/text input for keywords during generation so basically making predictions character by character about what the sentence is likely about. Then those extracted keywords could immediately query a RAG system or existing knowledge base to inject clean, reliable information directly into the context without interrupting the LLM’s inference flow. That way, instead of constantly doing separate tool calls (which cost extra tokens and latency) the model could handle everything in a single coherent prompt pipeline In general, I think time is still a badly handled component in agents. Most agents don’t really “understand” that their own generation takes time. Makes sense, since training data usually doesn’t encode that kind of hardware/runtime awareness. You can see weird effects sometimes like an agent creating a 10 second timeout before a command, even though the last tool use already happened more than 10 seconds ago before the timeout was even called. Really curious to see where all of this goes. The wheel has already been reinvented, we just have to learn how to use it properly.

its nice to see the industry converging on my work.