RAG isn’t broken. Your retrieval probably is. A lot of RAG applications fail for a simple reason:...RAG isn’t broken. Your retrieval probably is. A lot of RAG applications fail for a simple reason:...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Youssef's avatar
Youssef Talaat
• Sep 3
RAG isn’t broken. Your retrieval probably is.
A lot of RAG applications fail for a simple reason:
The LLM never receives the right information.
You can use the latest LLM. You can use a powerful embedding model. You can build a beautiful chatbot.
But if your retrieval returns irrelevant context…
your answer will probably be irrelevant too.
While building a RAG system, I focused heavily on the retrieval layer:
🔹 Dense vector search 🔹 BM25 keyword retrieval 🔹 Hybrid retrieval 🔹 Reciprocal Rank Fusion 🔹 Query expansion 🔹 Chunking optimization 🔹 Relevance evaluation
The lesson?
RAG quality starts before the LLM generates a single token.
A better architecture is:
Question → Retrieve → Rank → Select Evidence → Generate → Cite
Not:
Question → LLM → Hope for the best.
This is why I enjoy building RAG systems, AI knowledge assistants, and LLM applications — the interesting engineering happens between the question and the answer.
If you're building a RAG application, what are you optimizing first: retrieval or generation?
generativeainlprag
Post image
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started