-
🧵 I've been experimenting with #RAG (Retrieval-Augmented Generation) to query my codebase, hoping to boost a local #LLM like #Llama or #CodeGemma with code knowledge. But honestly? The results have been underwhelming...Permalink On twitter.com
Mood +5 🙂
-
When I dive into the output from a #vectorstore query, it's clear why: the results usually don't capture the structure of the code—just fragments of what's "close" to what the LLM thinks I'm asking about. 🤔Permalink On twitter.com
Mood +1 🙂
-
#RAG might be great for unstructured text (like PDFs), but when it comes to code? I'm starting to think it needs a different approach. Code is inherently structured—especially with OOP (classes, attributes, methods, etc.).Permalink On twitter.com
Mood +5 🙂
-
So here’s a thought: What if we used a structured system instead? Imagine a relational database storing all the details of your codebase, and the LLM queries it directly via a specialized interface. 🗂️On twitter.com
Mood 0
-
This way, the LLM could tap into the full class structure, answer high-level questions, and truly "understand" the code. When I paste an entire class into the AI's context, it gets it perfectly. So why not make this process programmatic?Permalink On twitter.com
Mood +3 🙂
-
I've looked around for solutions like this but haven’t found much... Has anyone out there stumbled upon something similar? Would love to hear your thoughts! 💡Permalink On twitter.com
Mood +9 🙂