RAG isn't the bottleneck in coding agents, it is the way you parse and chunk the codebase. Couldn't you just create chunks explaining the code base, run multiple llms to extract the aspects of the code base a human is looking at and reach quality while being price effective!?
He already did — the key is to equip the agent with a tool that can build a mental model of the codebase.Some latest RAG systems actually do this already: they create a structural map of knowledge, then let the model walk that map to retrieve relevant context. you should let the agent explore the codebase like a developer would, not just search it like a database.
Try pageindex https://github.com/VectifyAI/PageIndex, which does rag without any vectors and more like human
RAG isn't the bottleneck in coding agents, it is the way you parse and chunk the codebase. Couldn't you just create chunks explaining the code base, run multiple llms to extract the aspects of the code base a human is looking at and reach quality while being price effective!?
Said everything but explain what to use properly instead of RAG. 4/10
He already did — the key is to equip the agent with a tool that can build a mental model of the codebase.Some latest RAG systems actually do this already: they create a structural map of knowledge, then let the model walk that map to retrieve relevant context. you should let the agent explore the codebase like a developer would, not just search it like a database.
Soooo… knowledge graphs?
Use LLMs directly (CAG) as current SOTA models support larger contexts lengths. Enough to fit an entire codebase(s)