Discussion on "How RAG Works, End to End: From Raw Documents to Grounded Answers"

ZyVOP · 2026-10-20T03:21:53.380Z

A language model knows a great deal, and none of it is about your company’s refund policy, last week’s incident report or the contract you signed on Tuesday. It also has a habit of answering confident

Framing RAG as a search system with a language model on the end is the clearest mental model of it. Most bad answers really are retrieval misses upstream, and walking the failure modes per stage is more useful than another architecture diagram.

Exactly. That was the main idea I wanted to get across. A bad answer does not automatically mean the model needs to be better. Sometimes the document was parsed incorrectly, the chunk lost important context, or retrieval never surfaced the right passage in the first place.

That is why I think separating retrieval quality from generation quality matters so much. Once you can identify whether the right chunk was retrieved, ranked correctly, and actually passed into the prompt, debugging becomes much more systematic. Otherwise, it is easy to spend time changing models when the real problem is somewhere earlier in the pipeline.

The paragraph-based chunker needs one additional size guard: a single paragraph can exceed max_tokens, because it is appended after the overflow check and is never split. That matters when the embedding endpoint truncates long inputs without rejecting them.

The stored chunk can then contain a relevant sentence that never contributed to its vector, making the failure look like weak semantic retrieval. I would add a long-paragraph fixture and assert the final token count using the embedding tokenizer, including any contextual prefix. Record truncation explicitly or reject oversized chunks before indexing, so the text you inspect is the text the embedding model actually saw.

Good catch. You're right — a long paragraph can slip past the size check, and if the embedding model truncates it, the stored text won't match what the model actually used to create the vector. Adding a test for that case, including the contextual prefix, makes sense. I'd rather catch oversized chunks before indexing than have them silently affect retrieval. Thanks for pointing this out.

Search Hashnode

How RAG Works, End to End: From Raw Documents to Grounded Answers

Responses4