Boris from the Claude Code team explains why they ditched RAG for agentic discovery. "It outperformed everything. By a lot"

It's just another form of RAG, imo this is playing a semantics game to confuse people new to the field. Imo it doesn't matter if the model decides what retrieval/search tool to use, if you're still augmenting the context window to ground the generation, it's RAG.

Good stuff, I need to listen this now. Here’s the @latentspacepod link, if anybody needs it.

@latentspacepod Yeah, high signal to noise Thanks to crystalpeaks25 for DMing me the clip

It’s almost like RAG was a stop gap measure to fix weak models

I’ve read your post and it seems like you mean that RAG = embedding-based retrieval. I think it’s a broader concept than that. In the clip, he says they instead of using external embedding indexing, they use local search tools to augment the model. He doesn’t say that they ditched RAG.

Did it augment its context with information it retrieved in order to generate better output? Then it’s RAG.

thats me in the clip! btw Boris is back to speak more about Claude Code at come with

It’s only rag if made by a human, apparently

the whole argument in this paper is that a senior coding engineer would look through all the imports, dependencies, function calls and just generally get a good feel of the codebase.... and this is why RAG doesn't work. but you can literally achieve this better by doing GraphRAG

Total pash victory