The Fixed-Pipeline Assumption

Where we left off

Lesson 1 showed the problem in words: a fixed pipeline retrieves unconditionally, even for questions that don't need it. This lesson builds that fixed pipeline for real, the exact shape naive_rag through corrective_rag all share, embed every fixture note, retrieve the top k for whatever question arrives, stuff it into the prompt, generate, then runs it against both of Lesson 1's questions to watch the waste happen directly instead of just being told about it.

The code, piece by piece

def fixed_pipeline_ask(query: str, store: list[dict]) -> str:
    retrieved = retrieve(query, store)
    ...

Nothing here is new, it's naive_rag Lesson 8's whole pipeline in one function, unchanged. The only thing this lesson adds is calling it on a question, the boiling-point one, where retrieval was never necessary, and watching what happens anyway: an embedding call for the query, a similarity search across every note, and a prompt stuffed with the closest-matching (but still irrelevant) fixture content, all before Gemini ever gets asked anything.

Why this matters

Gemini usually still answers the boiling-point question correctly, it's common enough knowledge that a few irrelevant paragraphs about a sourdough starter or a pedalboard don't derail it. But "usually still gets it right despite the noise" is not the same as "the noise was free." Every fixed-pipeline call pays for an embedding request and a similarity search regardless of whether the question needed either, and on a harder or more ambiguous general-knowledge question, irrelevant retrieved context can actively pull the model toward a wrong answer instead of a right one, not just waste time getting to a right one.